Yes, we know it's to do with ranks. As stated, we have another video coming up with more detail. Everything is rushed with a launch. Re: Speaking with Wendell, we'll hopefully be doing something next week! He seems to indicate 2x16GB is a sweet spot, so we'll talk with him more about that soon and hope to feature him in a video! Watch our AMD R5 5600X CPU review: th-cam.com/video/01EhbmJAW-k/w-d-xo.html Watch our AMD R9 5950X CPU review: th-cam.com/video/72AHENDeTEI/w-d-xo.html Watch our AMD R9 5900X CPU review: th-cam.com/video/utWSSlyabjc/w-d-xo.html Watch our AMD R7 5800X CPU review: th-cam.com/video/6x2BYNimNOU/w-d-xo.html
@Gamers Nexus 2 sticks of Dual rank memory will do the same, Tom's did nice review about that on zen 2. The problem is that it's sometimes hard to know if the memory you are buying is single or dual rank.
Hi guys, I love the channel, but have recently started losing my hearing gradually, this has made me notice that you don't have any annotations on your channel, please could you add these for your videos in future so I can continue to enjoy your channel?
In the mean time you can use IBM's Watson Speech to Text free for up to 500 minutes a month. My wife uses it to transcribe whole presentations. Results will vary based on the person's voice.
So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
Woah! Stefan! I put an entire tube of thermal grizzly in the socket just like you said! Thanks for the tip. You're totally right about those those two sticks per second. I can only see one.
4xDIMMs single rank will be dual rank and that gives a performance boost sometimes, this is because the memory controller access them as dual rank memory on two channels. Basically the controller is dual channel but you can have single, dual and quad ranks. Ranks allow you to increase the amount of memory. Channels, dual channel means two command and data busses. A memory rank is a set of DRAM chips connected to the same chip select, which are therefore accessed simultaneously. In practice all DRAM chips share all of the other command and control signals, and only the chip select pins for each rank are separate (the data pins are shared across ranks). Modern memory devices are organized into ranks and each rank is divided into multiple banks, which can be accessed in parallel provided that no collisions occur on either buses. The two buses can be used in parallel: a request of one requestor can use the command bus while a request of another requestor uses the data bus. However, no more than one request can use the command bus (or data bus) at the same time. So each single ranked memory module can be connected as a single rank at each dimm slot. This means each rank has it own circuit. Say you have a Dual Rank memory module. You can only access one rank at a time because they both share the same circuit. When you select one rank, you cant select the second. Now with dual rank memory you can send commands to one rank and read data from the other rank. If they have their own circuits. In addition, modern systems can have multiple memory channels (i.e. multiple command and data bus). Each channel can be treated independently or they could be interleaved together. This means with 4xDIMMs single ranked. One channel can access one rank (first two slots) and the second can access the second rank (slot three and four). This means you can read form one rank and write to the other rank. Or read from both ranks at the same time. You cant do this if the first two slots have dual ranked ram. This is because both ranks share the same circuit. You can select only one rank at a time on the memory module. The chip select is the own difference on the PCB. The data circuit goes to both ranks on the RAM PCB. This implies to access one, you become locked out of the other. Many dual rank kits have poor timings, like CL16 and CL18. With a dual rank kit, there are normally chips on each size of the PCB (While most often the number of sides used to carry RAM chips corresponded to the number of ranks, sometimes they did not). So to get around the poor times of dual rank kits, you can run for example two single rank kits (4xDIMMs). So all of your 4 x DIMM slots are full. With my motherboard that means 3600MT/s dual Rank CL14 with decent b-die memory. This nets me near to ~800 point more in time spy cpu. I hit 15k approx. with a 10900k. My latency increases from 42ns single rank to 47ns dual rank as expected. This is with a CR of 2, so this would be better on a better motherboard than mine. I am no RAM expert but that what I got from A Composable Worst Case Latency Analysis for Multi-Rank DRAM Devices under Open Row Policy white paper. Zheng PeiWu - Rodolfo Pellizzoni - Danlu Guo
@@neidhartmuller8804 Run dual channel by throwing 4 sticks of single rank memory sticks to your mobo, second option, 2 sticks of dual rank ram. Don’t be that fool running quad channel on mediocre cpu, unless you got threadripper
@@ClassicalTechnology So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
every time I watch one of these videos, I have to remind myself that my system performs just fine for my needs, and I have better things to spend the money on. But, but, those frames 🥺
Mine doesn't quite, but I'm waiting it out, anyway. I've been going through my 3DS, and Steam/GoG retro and indie, backlogs, instead of getting a new PC 😁.
@@GamersNexus Waiting for you next video explaining every one about Dual Rank and Single Rank memory affect on Ryzen, this is what you found in this video - dual rank 2*16GB gonna perform the same as 4*8gb. Also maybe you can ask why memory vendor's don't list this info on the product page and mostly you need to check MB QVL list to see if the memory is SR or DR.
@@hwgeek86 it’s honestly strange to me that Steve is making this sound like new behavior. This exact same behavior is seen in Zen 2 where having two dual rank sticks or four single rank sticks gives a similar boost to performance because of improved memory latency. Honestly disappointing to see Steve hype this as a new thing when it’s been very well documented
@@syenosis Interleaving usually doesn't provide this much gains though. Mostly it was 2-3%. However sometimes you could achieve tighter timings on 2 total ranks than 4 thus out performing the interleaving. (depending on cpu/dimms of course). But this radical of performance difference means 2 total ranks are now dead.
@@syenosis I was also surprised to see this video, it should not be a news for him, it's old news: tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html
Hey GN. It's been a busy couple of months with several launches and you're work is excellent. I revisited this video and there clearly is a gain on 4 RAM sticks vs 2 however, reading into memory overclocking recently, I found out that it could also be a characteristic of the motherboard and it's memory trace layout/topology rather than. Exploring this subject would help clarify (although it is complex) the decision making for RAM sticks (frequency, latency, true first word latency & quantity of sticks) when people make their decisions in buying RAM to match their Zen3 chip/motherboard combinaison.
Jeeze I made it 15 seconds in and I couldn't listen. Like I'm sure I could figure out all the words he says, but the mumbling and slurring is hard to listen to.
I can't listen to him on anything other than 2.0 speed (or most YT'ers for that matter). I guess I'm acclimated. I also don't want to watch 45min-3hr videos.. lol
I'd love to see a comparison of a single CPU (say the 5800X or 5900X) and it running DDR in 2x8gb, 4x8gb, 2x16, 4x16, 2x32, and 4x32 (the max for most MBs) just to see what the differences might be (you could even toss in 1x16 and 1x32 if bored :). You might have to standardize to a CL16, but I'm curious if going higher in total RAM and its configuration has any impact (positive or negative).
Ryzen has always been sensitive to memory ranking. It's definitely something that isn't widely advertised on memory spec pages and should be. I'd like to see two sticks of 16 GB compared to four sticks of 8 gb with everything staying constant. On previous Ryzen CPUs they're handled exactly the same where four sticks of 16 GB will cause a severe performance degradation as they're uncapable of handling high clocks.
4 x 16 gb of dual rank b-die at cl 14 3200 has no performance degradation over 2x16 GB of the same ram with my 3950x running at XMP settings. The memory benchmarks are the same before and after the upgrade to 64 gb. This will limit the maximum clock speed but does not cause any problems. If anything, this is actually the best high density config for the Zen 2 series as it can be used at XMP without tweaks unlike the 3600 CL 16 b die requires as it puts too much load on the memory controller. I look forward to upgrading to the 5950x while carrying over the same ram as I'm very pleased with the performance of stock, always being amongst the top scores for my hardware (running at stock). Higher ram speed is not worth the hassle for me.
Easy to not sleep when getting new hardware, especially so close to the year-end Holiday season makes you feel young again... like a kid in a candy store...... like a kid in a well-stocked computer hardware store.
@@eeeealmo To an extent. In past generations, it was better to run fewer sticks in order to achieve more stable frequencies and timings -- particularly on Zen 1, where it was tough to stabilize past 3000MHz initially.
Exactly my thoughts. Each stick has a rank prolly there's a difference with 4v2. With a max upto 4 ranks per channel would be really interesting. But the timings being looser could be a reduction in performance without more volts and tightening timings.
@@ISO-ex2lg My 9900k also performed better with 4 SR vs 2 SR. You do have to ensure rank interleaving is turned on in the BIOS. It's not just AMD from my experience.
I'm confused because some say 2 dual rank sticks is even better than 4 single rank sticks. So right now I'm torn between Trident Z Neo 3600MHz CL14 4x8GB SR vs Trident Z Neo 3200MHz CL14 2x16GB DR. Which peforms better on a 5600x? In other words, Dual rank vs +400mhz, which is more important?
@@grzadzr Most 2x16GB kits are dual rank. There are some exceptions like the new high speed 2x16 kits from Crucial. The absolute best 2x16GB kits you can buy are the new 4266C17 kits from GSkill. On Ryzen, you would have to run them below XMP frequency though, as the infinity fabric can't go that high.
Im not native english speaker, never leave my country. Happy to find out that even smoe of native ones said they cant keep up with his speed. Cuz I can barely keep up.
Whether we're talking single- or dual-rank DIMMS, using four sticks has some advantages of using two: Populating all four slots ensures you have the sticks in the right slots w/o reading the manual, and it looks cool.
Buildzoid has a video (~ 4 weeks ago) that seems very related to this topic. This could to be from the difference in the memory rank, where 2 sticks is 1 rank per channel vs. 4 sticks being 2 ranks per channel. The memory controller can interleave the allocated memory which can allow 1 rank to be read while the other fetches some requested data. If this is the same as what you've observed, then I believe this is also present in zen 2. Not sure if intel also has this feature. Edit: this would also explain why Wendell suggests 2x16GB to be the sweet spot since most (all?) 16 GB memory sticks are just 2 ranks of memory on 1 stick and are essentially the same as 2x8GB as far as the memory controller is concerned (OC stability concerns aside).
Building and selling high performance servers we have seen this difference for several years. As dram chip sizes got larger, single rank dimm became more common and a memory channel with only one dram rank performs 10-20% worse than configs where there are two or more ranks on each channel. GN should see if they can find 2 dimm set of 16gb dimm a with two ranks on each dimm (16 or 18 chips on the pcb with 8 or 9 on each side) and see if they get the same performance as 4 sticks
Hi Steve. It's not the number of sticks that matters. It's the number of memory ranks per channel. Buildzoid has done a video on this called "Testing XMP settings in dual and single rank". Because of interleaving between the ranks the memory controller can talk to the 2nd rank, while the 1st one is "busy". This also has the effect that the more ranks you add per channel frequency starts to matter more and timings affect performace less. Sorry for my bad English.
In theory running quad rank 4x16GB(same timings) should give another bump in performance. If the memory controller can handle it. Also this is how servers get away with using "bad" JDEC timings on their ecc memory. Servers simply have a ton of memory ranks per channel.
He knew about it(that’s why they made that RAM choice for testing). The video is about the fact that is has quite a significant difference in case of Zen 3, which was unknown.
So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
This is crazy considering my 2700x X470 can't handle 4 sticks at 3200 (XMP) but can with 2 sticks. It's good to see 4 stick configurations doing better.
Long story short: it's not about 2 sticks is slower than 4 sticks on Zen 3. Or 2x16 being faster than 2x8. It's about 2 single-rank sticks (taken from 4x8 kit) being slower than 2 dedicated double-rank sticks. So if buying 2 sticks, make sure they are double rank.
24:30 So you had the correct answer all along, but for some reason did a whole bunch of testing and presented it as some kind of mystery that's totally new with Zen 3. It isn't new at all. Zen 3 just shows bigger differences in some applications. Those G.SKILL TridentZ sticks have a single rank. Use two of them, and you get one rank per channel. Use four of them, and you get two ranks per channel. That in turn gets you rank interleaving, which increases performance, in basically the same way that using multiple channels gives you channel interleaving, which increases performance. All bog standard stuff with computer memory. Using two sticks of dual-rank memory will probably provide slightly better results, because it's only one DIMM per channel, but still two ranks per channel. More than two ranks per channel might allow a very slight increase in performance, but the achievable memory speed also goes down as the number of ranks per channel goes up, so the optimal choice is likely going to be higher clocks with just two ranks per channel.
I think the difference, here, is that Ryzen is memory hungry enough that it actually makes a difference outside of very specific tasks. I could be wrong, though.
The bigger difference is what's new. Previously, you could make up the difference with faster 1R kits, and that was cheaper, too, with very fast kits going beyond the interleaving improvements.
"Games REALLY don't need it." Cities: Skylines would like a word. My savegames with mods was reaching 14gb used, which prompted me to upgrading my RAM to 32gb. I went with 4x8gb, which is why I'm watching this.
This is memory interweaving. Dual rank dual channel. 2x16 gb would be the same mostly. It helps on all CPUs by a small bit. It seems Zen 3 likes it more.
@@tevlar im going to buy same kit from newegg was doing research on which one had samsung capacitors and its the one you mentioned, you should prob get another kit for 64gb quad channel XD I know I'm going to
3600 / 16 = 225 3800 / 18 = 211,11 3200 / 14 = 228.57 That's why to me, 3200CL14 is faster here, against the rest. I am looking for any F4-3600C14Q-64Gxxx kit right now 4x16GB, 3600MHZ, CL14-15-15-35 1.45V. Wonder if that will beat all these
With a 5800x first had 2x16GB 3200MHz single rank CL16 memory. Sold that and purchased 4x8GB 3600MHz single rank CL18 and my Timespy cpu score increased 8-10% (from 11300 to 12300) !
I know I'm a year late but all 16gb sticks are dual rank. There are not any single rank 16gb sticks on the market. Your score was because of the speed, nothing else. You went from quad rank to quad rank
I cant wait to see these test done on the gen 3 cpus with the new radeon gpu like the 6800xt 😳 , can't wait to see that full team red config with the 4x8 1:1:1 kit trident z!👑
Only have 2x8GB in my new rig with 5800X - was contemplating getting another 2x8, or hold out til proper update down the line, so this vid has perfect timing
Depends what resolution your running. At anything above 1080p the returns diminished very quickly. Went from 1080p to 1440p to ultra wide and now the CPU is barely ever the bottleneck.
@@h4ckantrieb950 the Ram for normal Desktop CPUs comes in SR (single rank) and DR (dual rank) Modules and DR Modules offer more performance. 4* SR DDR4 = DR in DualChannel
Hi guys, I just wanted to point out that using 4 sticks of ram automatically means you're using dual rank. Dual Rank should _always_ give a performance boost. (not by 10% but close if you're cpu limited) Please consinder for further testing: 2x16gb signle rank sticks vs 2x16 dual rank sticks vs 4x8 gb single rank sticks (in dual rank configuration) I don't think you have to test with different speeds or timings.
Yes, for comparability, timing should stay the same. But might be complicated to get dual ranks of same timings. I always use dual rank memory for my notebooks, the intel integrated graphics benefit even more from it.
@_________ Yeah, maybe he misspoke of got it wrong from Wendell. Steve never explained SR and DR at all in the video and I think 95 %of the people watching this don't really know about it. Anyway Wendell should enlighten us soon.
I have MSI B450 Tomahawk MAX. I had 8x4 DDR4 Kits and they were highly unstable even at 2133. So now I have 16x2 (32GB) dual channel. This is my message to everyone who owns MSI B450 Tomahawk MAX, get yourself dual channel memory and make your life easy.
Dude, never use 4 memories in a b450/b550/b650, those cards only have 2 efficient lanes, that means only 2 memories can run fast and the other two are standard. You have to go for an X570/X670, etc to use 4 clubs.
Curious if there's a meaningful difference when dual rank is taken into consideration. ie: 4x8gb SR (as tested) vs 2x16GB DR. I personally have the latter in 3200CL14 and since you mentioned interleaving there may not be much of a performance difference between the above two configurations.
So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
@@Traumatree Not if electrical interference from the empty slots reduces your overclock. 4 single-rank sticks on a dual-channel platform are effectively dual-rank. So, if you're going dual-rank, you should prefer to populate all slots.
WHEW IM GLAD I WATCHED RIGHT TO THE END. I almost added two more 16gb sticks to my list of things to go along with the ryzen 5900x I plan on purchasing. My 32gb 2x16 3400mhz im using with my 3700x should be fine... Wow thought I was gonna end up with 64gb of ram to get the max performance out of the new CPU.
64gb is a lot. I would opt to get 32gb at higher clocks. I'm running an old single rank kits of trident z 3866s at 1.5v and at 3733 16 16 20 38. 16gb total. Make sure to time them right. Mine get 45.2ns reads pretty happy about it for kits I got 5 years ago. The upside with 64gb is if you do compute workloads though so if you can go for it.. they look good but I'm one of the unfortunate people out of work right now with no compensation.
Considering that RAM price is low enough to make 64GB RAM something not so extravagant, could you test or give your opinion on 2x32 vs 4x16 ? Based on the mention of Wendell saying 2x16GB is the sweet spot, it seems we can't necessarily assume that filling 4 slots is always better than using 2 slots (for the same total capacity). Also, would X570 vs B550 make a difference ?
@Matthew Yeah an extremely clickbait and intentionally misleading title, I expect better of GN. This 4 stick nonsense is going to be a myth that is now going to be spread around for the entirety of Ryzen 5000's existence.
@@gromitnwallace5281 here's the thing, though, think about all the people out there with 2x8. They can literally just buy another 2x8 of the same stuff and raise their performance like 5-10%. THAT is the takeaway imo And honestly, like $100 for that amount of improvement on what is probably a $1000 rig is not even unreasonable from a price/performance standpoint
@@chapstickbomber Oh the information is definitely useful, but GN presented it in such a way that both the title and first 20+ minutes of the video are misleading to everyone, until you watch to the very end to see the caveat 'oh yeah this actually only applies to 8x4 vs 8x2'
Problematic Video. Only those who actually watch the whole 27 minutes, will hear that " 2x16gb sticks performs best of all ". Others who usually skip to the benchmarks, will now believe that as a general rule, using 4 sticks, no matter the capacity, boosts your performance.
Damn, now I'm happy that I bought a 2 kits of 2x8gb (by mistake), the 5000 series will be a worth while upgrade in the future (going from 2700x). Thanks Guys for the video ! Keep it up
I'm at work, so I can't watch the video with sound fully, but I've been researching more about the whole 2 sticks VS 4 with the 5900x since recently after stumbling on a video testing that and 4 sticks being only a bit better in certain instances.... And recently trying to build a PC for a friend and thinking of selling him my RAM to get 2 sticks for myself (would be discounted and he does video editing) and get those 2 sticks to have some RGB lighting. ..... And also reading more and more about how pretty much Threadripper boards can use quad channel and dual channels can have trouble with 4 sticks........ Yeah, it's not easy.
After seeing this, I added a matching kit of GSkill Ripjaws V 3200mhz to my rig and saw the Mersene/ Prime95 iterations drop from 30ms to 17-18ms with my Ryzen 5 3600x at stock 3.8ghz. This knocked 8-9 days off my workload. haha GG Steve xD
Ordering another 2x8gb of G skill Trident Z royal 3600mhz right now for my 5600X. Planned on upgrading to 32gb eventually anyway so I might as well now. GamerNexus literally builds my PC lmao.
This could be myths or rumors but I've read that silicone differences can cause issues between ram and you should only use a set kit of ram instead of buying 2 seperate 8 gig sticks or something
@@dragonwithak Sounds like a myth to me because every single build I've ever done I always start out with 8 or 16gb and upgrade later down the line to another set. Never had any issues. No difference between me buying 4x8gb 3600mhz set of G skill royal or 2 2x8gb 3600mhz set of G skill royal. As long as timing, speeds, and brands are the same there shouldn't be any issue. But who knows just because I never had an issue doesn't mean some people could.
@@MoizCOUK Bought these a year ago lmao returning is outa the picture, I have never had an issue upgrading to another set ever in any of my previous builds. G skill has always been reliable for me in that sense. Regardless it's not a big deal at all. Worst comes to worse if I do run into an issue I'll just return the set I recently purchased and buy a set of 4. Easy peasy.
I thought I had RAM nailed for my next build (will be done within a month) . Gaming is only a little important, I want to run DaVinci Resolve for editing and rendering 2k vids and later 4k vids. The plan was to start with 2 sticks of 32 GB each and then add 2 more later if needed. Looks to me like single ranked is totally out of the question for me. I just have to get Ryzen optimized RAM wherever I can get it, apparently, and not worry about this. Please don't tell me I don't need 32 GB for projects, I fill 16 GB just going overboard with Firefox tabs.........like when looking at the right side of TH-cam to pick stuff for later viewing.......for PC research.
Steve i think you made a mistake at 23:57, in his video with AnandTech's Ian they say that 2 sticks of dual-rank memory work better than 2 sticks of single rank, while youre saying that 2x16 GB single-rank works the best while quoting Level1Techs.
and then most mobo are daisy chain so 2x16gb is easier for memorycontroller to drive than 4x16gb or 4x8gb so this leads to higher clocks for memory with tighter timings.
Tech Jesus throws acronyms like they're used by everyone and their cat. Me who has no idea what they mean, I'm wondering why I'm still watching this. 😂
You're essentially testing one vs two ranks per channel. This performance bump is expected (more or less). Just ask Buildzoid, he'll provide an explanation.
2x16 in most sticks 4x8 as your 16GB stick will have 2x8GB one 8GB on each side. So essentially 2x16GB = 4x8GB www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html Already known in zen2
This is "Rank Interleaving" for you :P While one rank in a channel is busy refreshing or activating rows or such a second independent rank can be addressed by the IMC. The (lengthy) top comment on this somewhat explains this, but gets the channel width wrong: The Data-width on one memory channel is 64 Bit, so with 2 IMCs you get 128 bits of that, regardless of rank count. One would select a different rank via the CS-Bit ("Chip Select"), which is not a data- but a control-pin. Take a look at Micron's technical documents regarding their UDIMMs (and look for RDIMMs and LRDIMMs while you're at it!). For weird Edge-cases you can even have multiple ranks on one Memory IC (usually that would be used in 3D-Stacked memory).
Really interesting. Thanks for the info Steve and the rest of GN. Looking forward to watching a few more tests on memory config, but it's looking like 4 x CL16 at 3600MHz is the sweet spot again? Wonder how speed vs latency pans out this time too!
@@davidfoote696 seeing as most people learn about pc's to build their own pc, now exactly how they work, I don't see how your reasoning makes any sense. people learn how ram benefits your pc, whether it be about it being the quantity of ram or speed.
I’ve had 2x16 3200c16 forever and just scored another 2x16 3200c16 to make a total of 4x16…this video makes me happy because it’s also the trident z rgb ram. Looks like I’m getting at least 5% gains with my 5900x at lower resolutions…I usually run 1080p, sometimes 1440p so yeah. I’m stoked for gains. I mainly just wanted 64gz for more multitasking, but this is a welcome addition. Great video, thanks for all the testing..👏🏼
how is it running? rule of thumb is usually to not just buy 2 sticks that didn't come off of the same manufacturing line, but to sell the 2 you have and buy 4 that did, so im just curious
@@JaxDagger I have the same configuration but I bought two packages separately G.Skill RIPJAWS V 3600mhz cl16,19,19,39 F43600C16D-32GVKC I put them at 3200Mhz, so far no blue screen and I play warzone 2.0 for hours with this seasonal heat and without any failure
It appears that Zen 3 is just capable of taking a significant performance advantage out of rank interleaving. All previous gen Ryzen and Core did not benefit from interleaving well. Watch Buildzoid's video on memory ranks for more detail.
dual rank ALWAYS had a per clock advantage. the main problem is that in many scenarios it came with a massive clock penalty that negated the per clock advantage.
I use 4*8 dimms on Taichi x370 with r5 2600x and I saw more stable performance in games even if I lost 4 ms. Latency from 64ns to 68ns as I running slower cos are mixed b-die hynix AFR
Really interesting find, thank you guys a lot! I am very curious to see if/how this influences the new AMD "Smart Access Memory" feature when the Radeon 6000 cards arrive for benchmarks.
Did you ever get around to doing a follow up to this? Found this really interesting. Also building a new pc and would like to know if 2x16 being the sweet spot produced the same results as 4x8. On another note just bought myself a Gamers Nexus mod mat and can’t wait to use it!
Hey man awesome channel, keep up the good work. Thought you might get a kick out of this.... Running my i5 3570k @ 5.1 for the last few years now, it's held up over the years, but every now and then needs a bump in voltage. The VERY FIRST WEEK I had it together, I had it gaming at a friends house, and something caught the CPU fan and blocked it from spinning, the chip sat throttling for about 3 hours above boiling point, great way to break in a new (at the time) chip.
As a hardware engineer having worked several years on the development of PCIe controllers, I don't see any reason for that to affect the performance of SAM. The Smart Access memory basically enables the CPU to see directly more GPU memory using a BAR reconfiguration that enables it to show more than 256MB through PCIe. Usually the PCIe BARs have to be that small because it's still 32b legacy during boot up. The gpu memory is then accessed through this bar window plus a memory bank register that is written whenever the upper bits of the address need to be modified. So if the memory can directly be accessed without needing to modify the upper address separately it may improve the access time, hence a slight improvement in performance when doing this. This has nothing to do with the cpu memory so changing the CPU memory configuration won't affect this.
As i recall with RAM manufacturers, the reaosn they sell sticks as a "quad channel kit" is because they are all tested together in 1 system to ensure that each stick is compatable with the settings being advertised. If you were to get 2 "dual channel" kits and put them in your system, there might be slight inconsistencies that makes them unstable to keep the same frequiencies and timings advertised, as they were not tested together as a 4. Just osmehting to consider. Maybe a video to test this?
I only have a single data point, but I bought two 2x8gb kits for 4x8gb and they work fine. Based on what I read, as long as the timings (the XMP profile) and the manufacturer (samsung/hynix/etc) are the same there's very little risk with mixing two 2x kits together. And if you're buying both kits at the same time from the same place, it's pretty much a guarantee that they'll play nicely with each other.
This video has changed my life....I saw this video in morning, bought another 16GB stick and my games are getting improved fps. 144+ fps in doom eternal with RTX ON, previously I was struggling to keep 100 FPS. 100+ fps with DF optimized settings in cyber punk...god I missed so much frames in past 3 months. (I have 3060 Ti, i9 10850k@stock)
Dual-rank interleaving increases performance. 4 single-rank DIMMs run in dual-rank mode on a dual-channel CPU. You'll see the same improvement with 2 dual-rank 16 GB DIMMs on a dual-channel platform, but the remaining empty slots cause electrical interference and reduce your overclock. Similarly, you want to run 8 single-rank or 4 dual-rank DIMMs on a quad-channel platform.
@@eazen Yeah, I also find that pretty wierd. You ususally dont get much interference from empty slots compared to actual components with electricity flowing through them being there
@@eazen special boards dedicated to RAM OC usually don't have 4 slots, only 2, same goes with mini-ITX boards that are usually better memory overclockers than ATX. 4 slots require longer traces as they are spaced out, as well as more complicated layout thus it affects signal integrity, but this occurs only at very high overclocks usually. Doesn't mean that empty slots are creating interference, so dude just used the wrong wording, but in general he is right
Hi Steve, I think you misspoke at the end and meant to say that 2x16GB dual rank dims are a little better than 4x8GB single rank dims. 2x16GB single rank dims (ex. Crucial Ballistic Max 2x16GB) are only 128-bit, so you are not getting the benefit of interleaving that way. I would also be interested in seeing how 4x16GB dual rank dims perform in your future video because there is a good chance 512-bit would overload the memory controller and tank your achievable memory frequencies. I'm curious if the additional bandwidth of 512-bit at slower speed would be enough to overcome the FCLK penalty compared to 256-bit and higher speed. I suspect not... TL;DW for everyone else... For best performance you want to have a memory set up with 256-bit bandwidth meaning 2x dual rank dims(2x128-bit) or 4x single rank dims(4x64-bit). 2x single rank (ex. 2x8GB) can leave up to 10% performance on the table due to lack of interleaving.
Thanks, this helps with my memory purchasing plans going forward for a 5600X when stock is replenished. Hopefully you will consider using the 8x2 and 8x4 configs with testing the 6800/6900 and RTX 3000. I appreciate and love all you do and share with us!
A really impressive video. Steve and GN showed today that how reliable their benchmarking methodology is. The attention to detail from you guys is off the charts. I have been a sub for 2-3 years. I have subbed for life. :)
Whenever I ask this question of 2 vs. 4 sticks, most of the time tech reviewers always tell me to go with 2 sticks for (2) reasons...(1). Running (4) sticks puts a much higher workload on the memory controller of the processor, and generates more heat. (2). You can buy (2) sticks with lower lower timing settings as compared to 4-stick kits, giving better performance. Now its looks like for the the Zen 3 that this isn't true, and 4-stick kits are better?
Explanation is simple as that : 4x8Gb memory single rank double channel works as 2x16Gb dual channel double rank memory . And that means faster, because there are more 64 adresable units seen (double) and the memory controller can access the 64 units in parralel (for the dual rank memory). And this is the same for AMD or Intel CPU's (not only Zen 3 ...)
I was getting a 32 with 16x2 but after someone reviewed my part selection for my Gaming/Streaming Single PC Build. They recommended I got with 32 with 8x4 because they said I would get increased performance if all 4 slots are filled rather than just two. I also increased my boot up storage from 500 GB to 1TB.
Conclusion: Don´t compare different sites benchmarks, focus on the % difference between the products you want to compare for each reviewer. Also, 4x8 ram sticks has better performance than 2x8 but not because of memory capacity but because of number of ranks, so 2x16 dual rank might be even better. Great job and great finding, as always!
For some reason, even though I know exactly what Steve is saying during these benchmarks, my brain just can't keep up with the actual words, but somehow I still understand everything. It's like the words just bypass the conscious part of my brain. It's a trip!
Props to bz for his devotion to not editing though, even of he knows he's sending to GN that will edit. I imagine there's comedy gold in some fast talking Steve cuts... Maybe they're on patreon or something
Extremely informative as always. Wish the presenter would slow down a bit, your channel already does an excellent job of 'chaptering' the content to make it easy to find relevant information, there is simply no need to speak at 100 miles a minute. I could be in the minority, but I think I would rather have the video be 50-60 minutes instead of having to replay a segment 3, 4 or 5 times to follow what you're saying.
The probable reason is that even on same channel, with two memory module, memory controller can send a precharge commands to one stick, then execute memory transfer on the other, and by the time the transfer is finished, be able to do other commands on the first stick. Similarly, it probably utilizes better the burst transfer, by interleaving burst commands between two sticks. Very smart. It improves effective bandwidth and latency hiding. I don't think it is Ryzen 5000 specific, and it is likely similar tricks are done on previous Zen CPUs. It should be rechecked. The longer load / store buffers on Zen 3, maybe make it more visible, but it should be similarly visible on previous Zen CPUs with multi threaded workloads that use memory a lot.
Haven't finished yet but I assume it's rank interleaving being especially beneficial to the latest architecture? So shouldn't dual rank sticks work just as well?
@@dralord1307 Interesting, on the other hand four sticks can have poor compatibility with certain memory layouts which will hurt signal integrity hence speed and timings. Either way having finished the video it seems Steve's on the right track, hope he talks to buildzoid as he and HUB both made videos about this behavior on zen2
@@teddygoboom1 if you really want to understand I suggest reading this, www.anandtech.com/show/16214/amd-zen-3-ryzen-deep-dive-review-5950x-5900x-5800x-and-5700x-tested
@@teddygoboom1 4 sticks and stability was an issue all the way back in the Vishera/DDR3/FX CPU days. I'm not so sure dual ranked modules aren't also troublemakers as well.
I ran a Timespy benchmark with two sticks of G Skill Aegis 3200mhz, 8gb, 16-18-18-38, single-rank memory sticks installed. The Ryzen 5 5600x was at stock clock with increased boost capabilities through Asus PBO, and saw a peak of 4.85ghz during testing on all six cores at one time or another (no specific graphing available unfortunately). I then hopped on amazon and ordered two more identical sticks of memory. Once it arrived, I popped it in, checked that everything was working, and ran the test again with four ranks of memory. I saw a modest graphics score improvement but a whopping 8.3% spike in CPU score. This wasn't a perfect test. For one thing, windows updated between the time of the first score and when my ram came in the mail. Still, it's very interesting to me and I thought it would be worth sharing, regardless of the imperfections in the methodology. I may go ahead and pull the sticks and do a back-to-back retest on the same windows version to verify this result.
@@mrdwilkster Firstly, disclaimer: I’m no expert on RAM. My understanding is multi-rank memory has an advantage thanks to “rank interleaving”. Quoting Guru3D, “(Dual Rank) modules can often be a smidge faster thanks to a capability called “rank interleaving,” wherein the second memory rank can still perform work while the first is being refreshed for use.” www.guru3d.com/news-story/amd-blogs-on-ryzen-memory-performance.html Think of it as dual rank allows memory operations to occur in parallel or queue, like multithreading or hyper-threading is to CPUs. For memory this means there’s less penalty for high latency, versus single rank.
@@BurnsRubber any applications? Linpack and Geekbench are synthetics. Useful, but not in terms of what we should see elsewhere. The 5900X and 5950X should get more from added ranks, though, as they should be able to queue up up to double the L3 misses if the 5600X and 5800X.
@@laurelsporter Yes benchmarking codes doesn’t always translate into real world gains and agree the 12 and 16-core should see more benefit as they can become memory bandwitch starved. Although synthetic, the calculations Linpack and Geekbench do are relevant to scientific and engineering software I use. Real world software isn’t easy to benchmark. Was mostly pointing out that switching from SR --> DR and DR -> QR seems to allow me to get away with looser timings without loss in performance, and can even add some performance if you can mostly maintain low latency timings.
It will depend on if the 16 GB DIMMs are single rank (i.e. usually 8 DRAM chips total on the module) or dual rank (usually 16 DRAM chips).* Single rank means less capacitance on the control and address lines allowing for higher speed whilst dual rank allows the memory controller to issue more memory operations in parallel by interleaving requests to the different ranks. [*] I wrote usual as it depends on how many bits the individual DRAM chip width is as they can be 4, 8, or 16 per chip..
typically a quad channel kit will all be produced on the same prod line, exact same chip config for optimal compatibility in quad channel. You can get in sub optimal state with separate kits that may have not have exact same chips/chip configs. But yes as long as the ram sticks don't have too much variation, you can run four random sticks and you may still get quad channel stable.
Would love to see you have that conversation with Wendell on screen, I think that could be super informative for your viewers. Conversations between the most rigorous testing channel and The Janitor help me to think of questions I didn't even think to ask before seeing them. Thanks for all the great content and hard work!
Yes, we know it's to do with ranks. As stated, we have another video coming up with more detail. Everything is rushed with a launch. Re: Speaking with Wendell, we'll hopefully be doing something next week! He seems to indicate 2x16GB is a sweet spot, so we'll talk with him more about that soon and hope to feature him in a video!
Watch our AMD R5 5600X CPU review: th-cam.com/video/01EhbmJAW-k/w-d-xo.html
Watch our AMD R9 5950X CPU review: th-cam.com/video/72AHENDeTEI/w-d-xo.html
Watch our AMD R9 5900X CPU review: th-cam.com/video/utWSSlyabjc/w-d-xo.html
Watch our AMD R7 5800X CPU review: th-cam.com/video/6x2BYNimNOU/w-d-xo.html
Just got done watching the last one (5600X) and saw this pop up, was planning to just go to bed, but eh. I'll crank another one out.
Another great video. Cheers Steveo mate.
What was tLDW set to?
😉
@Gamers Nexus 2 sticks of Dual rank memory will do the same, Tom's did nice review about that on zen 2.
The problem is that it's sometimes hard to know if the memory you are buying is single or dual rank.
I mean, you're still getting their view right? Jokes on them when someone flames them xD
Everyone knows that peak performance is achieved with 1 stick of RAM and 3 Corsair RGB dummy sticks
I genuinely just had to Google that because I didn’t believe that was an actual thing.
Really huh?
@@brindaraka5899 yeah, im running that with my threadripper and 3090
But what timing and colour configuration should be applied to those spacer sticks? We need answers!
@@theevilmuppet 2
Hi guys,
I love the channel, but have recently started losing my hearing gradually, this has made me notice that you don't have any annotations on your channel, please could you add these for your videos in future so I can continue to enjoy your channel?
help this man
Ohh man, there is really no CC. Not good.
In the mean time you can use IBM's Watson Speech to Text free for up to 500 minutes a month. My wife uses it to transcribe whole presentations. Results will vary based on the person's voice.
ODD... this is a choice when setting up the video to youtube... just a checkbox (English by TH-cam (automatic))
like why not just allow autocaptions
It looks like it's NO SLEEP NOVEMBER for you guys.
Really appreciate all the content!.
Don't forget the few more wk amd gpu launch.
Im just in Normal Neck November
So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
Well said. Incredible efforts.
@@iamLODD plus their benchmark suit now has to change with AMD basically being all round better now.
Well, this explains why some reviewers had Zen 3 performing on average on par with Intel, whereas others had it performing 10% better.
Yep. Totally makes sense now.
they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
@@iamLODD I'm surprised AMD didn't brief press about this in their architecture day... Instead they just let outlets discover it for themselves.
@@iamLODD No one has to "redo" anything. It's all fine as long as the testing details what is used. What a bizarre overreaction.
@@georgelopez9411 Just wait until reviewers get to play with S.A.M.
The human eye can't even see more than 2 sticks per second.
Lol the difference between 100fps and 1,000,000fps is null the human is the limiting hardware
The verge pc building guy😂😂
😂😂😂😂😂 Does he know you made this account???
Woah! Stefan! I put an entire tube of thermal grizzly in the socket just like you said! Thanks for the tip. You're totally right about those those two sticks per second. I can only see one.
Stefan Etienne it's true, we got one!
Only Gamer's Nexus could make a 27 minute video on the difference between 2 sticks of ram and 4 sticks of ram.
what was his conclusion?- i play all vids at 2x and missed it.
@@randymarsh3432 just consider anything featuring Steve talking to already be playing at 2x already.
And that's why I watch their stuff. Comprehensive.
😂😂😂
And we looovveee it
When the big and experienced tech TH-camrs need tech support, they call Wendell and I think that's beautiful.
I have a selfie with Donald Trump - can we trade?
Nice seeing you around after Robocraft Baerentoeter
4xDIMMs single rank will be dual rank and that gives a performance boost sometimes, this is because the memory controller access them as dual rank memory on two channels. Basically the controller is dual channel but you can have single, dual and quad ranks. Ranks allow you to increase the amount of memory. Channels, dual channel means two command and data busses. A memory rank is a set of DRAM chips connected to the same chip select, which are therefore accessed simultaneously. In practice all DRAM chips share all of the other command and control signals, and only the chip select pins for each rank are separate (the data pins are shared across ranks).
Modern memory devices are organized into ranks and each rank is divided into multiple banks, which can be accessed in parallel provided that no collisions occur on either buses. The two buses can be used in parallel: a request of one requestor can use the command bus while a request of another requestor uses the data bus. However, no more than one request can use the command bus (or data bus) at the same time.
So each single ranked memory module can be connected as a single rank at each dimm slot. This means each rank has it own circuit. Say you have a Dual Rank memory module. You can only access one rank at a time because they both share the same circuit. When you select one rank, you cant select the second.
Now with dual rank memory you can send commands to one rank and read data from the other rank. If they have their own circuits.
In addition, modern systems can have multiple memory channels (i.e. multiple command and data bus). Each channel can be treated independently or they could be interleaved together. This means with 4xDIMMs single ranked. One channel can access one rank (first two slots) and the second can access the second rank (slot three and four). This means you can read form one rank and write to the other rank. Or read from both ranks at the same time. You cant do this if the first two slots have dual ranked ram. This is because both ranks share the same circuit. You can select only one rank at a time on the memory module. The chip select is the own difference on the PCB. The data circuit goes to both ranks on the RAM PCB. This implies to access one, you become locked out of the other.
Many dual rank kits have poor timings, like CL16 and CL18. With a dual rank kit, there are normally chips on each size of the PCB (While most often the number of sides used to carry RAM chips corresponded to the number of ranks, sometimes they did not). So to get around the poor times of dual rank kits, you can run for example two single rank kits (4xDIMMs). So all of your 4 x DIMM slots are full. With my motherboard that means 3600MT/s dual Rank CL14 with decent b-die memory. This nets me near to ~800 point more in time spy cpu. I hit 15k approx. with a 10900k. My latency increases from 42ns single rank to 47ns dual rank as expected. This is with a CR of 2, so this would be better on a better motherboard than mine.
I am no RAM expert but that what I got from A Composable Worst Case Latency Analysis for Multi-Rank DRAM Devices under Open Row Policy white paper. Zheng PeiWu - Rodolfo Pellizzoni - Danlu Guo
ummm tl:dr?
@@raptorhacker599 indeed.
@@GodKitty677 my man can you please summarize so my ape brain can comprehend this
@@neidhartmuller8804 Run dual channel by throwing 4 sticks of single rank memory sticks to your mobo, second option, 2 sticks of dual rank ram. Don’t be that fool running quad channel on mediocre cpu, unless you got threadripper
"I am no RAM expert".. If it talks like a ram expert, and it looks like a ram expert....
Fantastic testing, I was just thinking this - lots of benchmark differences in some cases, memory certainly a factor. Thank you!
Big fan of yours, its fun to see you commenting here!
@@alexandreparent5754 nice to see you here too!
@@ClassicalTechnology So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
All I'm thinking during this video is: If size doesn't matter, only the number of sticks, does 4x4gb run better than 2x8gb then?
@@shuffleducky7803 yes, weirdly yes
every time I watch one of these videos, I have to remind myself that my system performs just fine for my needs, and I have better things to spend the money on. But, but, those frames 🥺
Someday. 😉
Mine doesn't quite, but I'm waiting it out, anyway. I've been going through my 3DS, and Steam/GoG retro and indie, backlogs, instead of getting a new PC 😁.
I literally just went but another 2 sticks of ram just because of this video lmao
That's me buying a 5800x then coming home to look up videos about the 5800x3d .. Gotta tell myself to cut it out
I just wanna know if i can put 8x4 stick for 32gb, then i buy it... it look ugly if there's is empty slot..
Me on my 16gbs of ddr3: I like your words magic man
What? I can't hear you all the way down here on ddr2
I used to use a Notebook with DDR RAM
@@webdude15 You have to speak up, I can't hear you over my 64GB of RGB DDR4
@@Renegade0056 (whispering) damn, i cant speak up my wallet isnt full enough :(
@gmu_alum08 can't upgrade a laptop
Hey Steve I got my bar runner earlier today! I love it thank you!
Awesome to hear that!
@@GamersNexus Waiting for you next video explaining every one about Dual Rank and Single Rank memory affect on Ryzen, this is what you found in this video - dual rank 2*16GB gonna perform the same as 4*8gb.
Also maybe you can ask why memory vendor's don't list this info on the product page and mostly you need to check MB QVL list to see if the memory is SR or DR.
@@hwgeek86 it’s honestly strange to me that Steve is making this sound like new behavior. This exact same behavior is seen in Zen 2 where having two dual rank sticks or four single rank sticks gives a similar boost to performance because of improved memory latency. Honestly disappointing to see Steve hype this as a new thing when it’s been very well documented
@@syenosis Interleaving usually doesn't provide this much gains though. Mostly it was 2-3%. However sometimes you could achieve tighter timings on 2 total ranks than 4 thus out performing the interleaving. (depending on cpu/dimms of course). But this radical of performance difference means 2 total ranks are now dead.
@@syenosis I was also surprised to see this video, it should not be a news for him, it's old news: tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html
Steve : Finds out about memory behavior
AMD : It's not a bug, it's a feature.
hidden feature.. or easter egg
Well, it's weird, but way better than the other way around.
Probably both
@Cthulhu hahaha, goes well with your avatar
@@OneTwoThree225 this was for zen2 already. It has been known for a long time
Hey GN. It's been a busy couple of months with several launches and you're work is excellent. I revisited this video and there clearly is a gain on 4 RAM sticks vs 2 however, reading into memory overclocking recently, I found out that it could also be a characteristic of the motherboard and it's memory trace layout/topology rather than. Exploring this subject would help clarify (although it is complex) the decision making for RAM sticks (frequency, latency, true first word latency & quantity of sticks) when people make their decisions in buying RAM to match their Zen3 chip/motherboard combinaison.
If Steve ever wants to switch jobs, he can easily be an auctioneer.
Jeeze I made it 15 seconds in and I couldn't listen. Like I'm sure I could figure out all the words he says, but the mumbling and slurring is hard to listen to.
lol
Listen to Steve at 0.75 speed and he's my favorite drunk.
@@TheMikeyb86 that sounds like the best way to spend the rest of my afternoon.
I can't listen to him on anything other than 2.0 speed (or most YT'ers for that matter). I guess I'm acclimated. I also don't want to watch 45min-3hr videos.. lol
I was LITERALLY googling for this answer earlier today! Thank you Steve & co.!!
Same here
I'd love to see a comparison of a single CPU (say the 5800X or 5900X) and it running DDR in 2x8gb, 4x8gb, 2x16, 4x16, 2x32, and 4x32 (the max for most MBs) just to see what the differences might be (you could even toss in 1x16 and 1x32 if bored :). You might have to standardize to a CL16, but I'm curious if going higher in total RAM and its configuration has any impact (positive or negative).
Would love that also at cl14.
Ryzen has always been sensitive to memory ranking. It's definitely something that isn't widely advertised on memory spec pages and should be.
I'd like to see two sticks of 16 GB compared to four sticks of 8 gb with everything staying constant. On previous Ryzen CPUs they're handled exactly the same where four sticks of 16 GB will cause a severe performance degradation as they're uncapable of handling high clocks.
Watch Channel, Odin Hardware, he has done this on a triple side by side test of many games per video
4 x 16 gb of dual rank b-die at cl 14 3200 has no performance degradation over 2x16 GB of the same ram with my 3950x running at XMP settings. The memory benchmarks are the same before and after the upgrade to 64 gb. This will limit the maximum clock speed but does not cause any problems. If anything, this is actually the best high density config for the Zen 2 series as it can be used at XMP without tweaks unlike the 3600 CL 16 b die requires as it puts too much load on the memory controller.
I look forward to upgrading to the 5950x while carrying over the same ram as I'm very pleased with the performance of stock, always being amongst the top scores for my hardware (running at stock). Higher ram speed is not worth the hassle for me.
Steve are you a cyborg who doesn't sleep?
It’s a miracle which makes sense for Tech Jesus
Higher is better, more consistent is
This is clone #4, original Steve is currently on vacation.
@@TheMonthlyJack Actually, the original Steve died from exhaustion, but don't worry, he'll respawn in 3 days.
Easy to not sleep when getting new hardware, especially so close to the year-end Holiday season makes you feel young again... like a kid in a candy store...... like a kid in a well-stocked computer hardware store.
You guys, Wendell and Buildzoid. Best tech guys on youtube, seriously. Glad you see you get them involved in the show Gamers Nexus.
I’d like to see an in-depth look into the best memory config with a set amount. 1x16GB, 2x8GB, and 4x4GB
Hey dont forget 8x2GB!
1x16gb is already ruled out because that would only run single channel. You should have 2 ram sticks at least.
memory guy here. the more sticks the better, as it gives you more concurrency, and more memory banks to respond to requests (thereby reducing latency)
1x16 would not make any sense to run.
@@eeeealmo To an extent. In past generations, it was better to run fewer sticks in order to achieve more stable frequencies and timings -- particularly on Zen 1, where it was tough to stabilize past 3000MHz initially.
Try testing dual rank sticks! It's likely the performance will go up some more!
Exactly my thoughts. Each stick has a rank prolly there's a difference with 4v2. With a max upto 4 ranks per channel would be really interesting. But the timings being looser could be a reduction in performance without more volts and tightening timings.
@@ISO-ex2lg My 9900k also performed better with 4 SR vs 2 SR. You do have to ensure rank interleaving is turned on in the BIOS. It's not just AMD from my experience.
Can you name some RAM dual rank modules? I'm looking for some info and tests with them.
I'm confused because some say 2 dual rank sticks is even better than 4 single rank sticks. So right now I'm torn between Trident Z Neo 3600MHz CL14 4x8GB SR vs Trident Z Neo 3200MHz CL14 2x16GB DR. Which peforms better on a 5600x? In other words, Dual rank vs +400mhz, which is more important?
@@grzadzr Most 2x16GB kits are dual rank. There are some exceptions like the new high speed 2x16 kits from Crucial.
The absolute best 2x16GB kits you can buy are the new 4266C17 kits from GSkill. On Ryzen, you would have to run them below XMP frequency though, as the infinity fabric can't go that high.
Im not native english speaker, never leave my country. Happy to find out that even smoe of native ones said they cant keep up with his speed. Cuz I can barely keep up.
Whether we're talking single- or dual-rank DIMMS, using four sticks has some advantages of using two: Populating all four slots ensures you have the sticks in the right slots w/o reading the manual, and it looks cool.
Harder to overclock tho.
Monos have diagrams, at least mine have.
@@rammwurst360 i mean you probably have no business ocing ram if you can't remember which slots to put your sticks in...
Buildzoid has a video (~ 4 weeks ago) that seems very related to this topic. This could to be from the difference in the memory rank, where 2 sticks is 1 rank per channel vs. 4 sticks being 2 ranks per channel. The memory controller can interleave the allocated memory which can allow 1 rank to be read while the other fetches some requested data. If this is the same as what you've observed, then I believe this is also present in zen 2. Not sure if intel also has this feature.
Edit: this would also explain why Wendell suggests 2x16GB to be the sweet spot since most (all?) 16 GB memory sticks are just 2 ranks of memory on 1 stick and are essentially the same as 2x8GB as far as the memory controller is concerned (OC stability concerns aside).
Rank interleaving has been around for decades, it's nothing new.
@@Walrusbonzo I'm wondering why they posted it only now
Intel also has memory rank interleaving. It's been around since the original DDR.
its a dual channel motherboard in this vid right?
@@0cards0 yep
Building and selling high performance servers we have seen this difference for several years. As dram chip sizes got larger, single rank dimm became more common and a memory channel with only one dram rank performs 10-20% worse than configs where there are two or more ranks on each channel.
GN should see if they can find 2 dimm set of 16gb dimm a with two ranks on each dimm (16 or 18 chips on the pcb with 8 or 9 on each side) and see if they get the same performance as 4 sticks
Hi Steve.
It's not the number of sticks that matters. It's the number of memory ranks per channel.
Buildzoid has done a video on this called "Testing XMP settings in dual and single rank".
Because of interleaving between the ranks the memory controller can talk to the 2nd rank, while the 1st one is "busy".
This also has the effect that the more ranks you add per channel frequency starts to matter more and timings affect performace less.
Sorry for my bad English.
You explained it the best in any of the comments I've read so don't worry about your English
In theory running quad rank 4x16GB(same timings) should give another bump in performance. If the memory controller can handle it.
Also this is how servers get away with using "bad" JDEC timings on their ecc memory. Servers simply have a ton of memory ranks per channel.
@@yosharian Thanks.
I would think Steve knows about this. This must be different than that, because he claims the effect is greater on Zen3 than Zen2
He knew about it(that’s why they made that RAM choice for testing). The video is about the fact that is has quite a significant difference in case of Zen 3, which was unknown.
Zen 3: *Run faster with 4 sticks RAM*
RAM store stocks: *chuckle* I'm in danger
My wallet is in danger too
XD legend
@ Yes, it is cheaper to get 2x16.
Less latecy = Better = more expensive
So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
@ i have 2x8 now and I will upgrade to 4x8.
This is crazy considering my 2700x X470 can't handle 4 sticks at 3200 (XMP) but can with 2 sticks. It's good to see 4 stick configurations doing better.
Long story short: it's not about 2 sticks is slower than 4 sticks on Zen 3. Or 2x16 being faster than 2x8. It's about 2 single-rank sticks (taken from 4x8 kit) being slower than 2 dedicated double-rank sticks. So if buying 2 sticks, make sure they are double rank.
What is double rank?
@@lastfirst5863 I meant dual-rank.
@@reav3rtm I found out what that meant elsewhere, but I really had no idea what that even was.
24:30 So you had the correct answer all along, but for some reason did a whole bunch of testing and presented it as some kind of mystery that's totally new with Zen 3.
It isn't new at all. Zen 3 just shows bigger differences in some applications.
Those G.SKILL TridentZ sticks have a single rank. Use two of them, and you get one rank per channel. Use four of them, and you get two ranks per channel. That in turn gets you rank interleaving, which increases performance, in basically the same way that using multiple channels gives you channel interleaving, which increases performance. All bog standard stuff with computer memory.
Using two sticks of dual-rank memory will probably provide slightly better results, because it's only one DIMM per channel, but still two ranks per channel.
More than two ranks per channel might allow a very slight increase in performance, but the achievable memory speed also goes down as the number of ranks per channel goes up, so the optimal choice is likely going to be higher clocks with just two ranks per channel.
I think the difference, here, is that Ryzen is memory hungry enough that it actually makes a difference outside of very specific tasks. I could be wrong, though.
The video signifies the fact that not all channels tested with optimal configuration and that Ryzen 5000 seems to need DR even more than Zen 2.
The bigger difference is what's new. Previously, you could make up the difference with faster 1R kits, and that was cheaper, too, with very fast kits going beyond the interleaving improvements.
@Transistor Jump very likely no. 16gb memories are double rank
thanks for this comment! my heart jumped a bit as i have a 2x16gb tridentz neo kit (dual rank). i think i'm set for zen3
"Games REALLY don't need it."
Cities: Skylines would like a word. My savegames with mods was reaching 14gb used, which prompted me to upgrading my RAM to 32gb. I went with 4x8gb, which is why I'm watching this.
To be fair this is pretty much one of the only games that actually will blow through an entire 16GB.
Me here
This is memory interweaving. Dual rank dual channel. 2x16 gb would be the same mostly.
It helps on all CPUs by a small bit. It seems Zen 3 likes it more.
Thanks for the Info! as I'm currently running an ITX system.
So I shouldn't trade my 2x16gb for 4x8gb? I just traded my 4x8gb for the 2x16gb when I upgraded to 5900x haha
Dual rank was also good on older gen zen processor if you could get the frequency and trimming down.
@@juststeve6296 if its dual rank then no
Interleaving
Are you planning an updated "best ram for ryzen"?
@@tevlar they go hand in hand, probably just grab 3200cl16
Depends if you want to pump up the memory speed some or lower the latency
@@tevlar im going to buy same kit from newegg was doing research on which one had samsung capacitors and its the one you mentioned, you should prob get another kit for 64gb quad channel XD I know I'm going to
DO IT! I am burning to know how 3600CL14 and 3800CL15-16 would perform
3600 / 16 = 225
3800 / 18 = 211,11
3200 / 14 = 228.57
That's why to me, 3200CL14 is faster here, against the rest.
I am looking for any F4-3600C14Q-64Gxxx kit right now 4x16GB, 3600MHZ, CL14-15-15-35 1.45V.
Wonder if that will beat all these
With a 5800x first had 2x16GB 3200MHz single rank CL16 memory. Sold that and purchased 4x8GB 3600MHz single rank CL18 and my Timespy cpu score increased 8-10% (from 11300 to 12300) !
3200mhz to 3600mhz
I know I'm a year late but all 16gb sticks are dual rank. There are not any single rank 16gb sticks on the market. Your score was because of the speed, nothing else. You went from quad rank to quad rank
@@weavercs4014 16gb sticks can be single rank, also 3200cl16 is the same as 3600cl18, same nanoseconds
I cant wait to see these test done on the gen 3 cpus with the new radeon gpu like the 6800xt 😳 , can't wait to see that full team red config with the 4x8 1:1:1 kit trident z!👑
Only have 2x8GB in my new rig with 5800X - was contemplating getting another 2x8, or hold out til proper update down the line, so this vid has perfect timing
Que elegirás?
I remember reading something about that in a pc parts manual so not 100% surprised by your result.
A pleasure to watch you in action!
This is a interesting test. Maybe people will actually start using all four slots.
Depends what resolution your running. At anything above 1080p the returns diminished very quickly. Went from 1080p to 1440p to ultra wide and now the CPU is barely ever the bottleneck.
I had no idea people didn't use all 4. I mostly use my computer for production so I need the ram. My motto is if there's a hole, it must be plugged!
@@jakejohnson1378 guess I don't need to ask where you store your phone!
@@jakejohnson1378 i see you man of culture, here i'm thought dp already push the limit, my respect for you brother
@@ablet85 That only applies to gaming for 1440p, and not if you are running high FPS.
AMD:"we give a reason to fill all 4 slot memory in your motherboard!"
Wong. It‘s just 2*16gb sr and not 2*16gb dr ram..
so of course 4 sticks offer better performance.
@@iDoe84 what?
@@h4ckantrieb950 the Ram for normal Desktop CPUs comes in SR (single rank) and DR (dual rank) Modules and DR Modules offer more performance.
4* SR DDR4 = DR in DualChannel
@@iDoe84 I think you are wrong. Check the comment of Mirsad Redzovic
"the more you buy, the more you save"
Thanks for giving me an excuse for 2 more sticks. Seriously, was on the fence
Bless you for doing this, I've been wondering this for my new build.
What you got for new build?
@@OGNubbs I wanted 64gb total, ended up getting a crucial Ballistix 2x32gb 3600mhz cas 16 kit
@@stevetb7777 oh that's a great set you got! Congrats
Hi guys, I just wanted to point out that using 4 sticks of ram automatically means you're using dual rank. Dual Rank should _always_ give a performance boost. (not by 10% but close if you're cpu limited)
Please consinder for further testing:
2x16gb signle rank sticks vs 2x16 dual rank sticks vs 4x8 gb single rank sticks (in dual rank configuration)
I don't think you have to test with different speeds or timings.
Yes, for comparability, timing should stay the same. But might be complicated to get dual ranks of same timings. I always use dual rank memory for my notebooks, the intel integrated graphics benefit even more from it.
@_________ Yeah, maybe he misspoke of got it wrong from Wendell. Steve never explained SR and DR at all in the video and I think 95 %of the people watching this don't really know about it. Anyway Wendell should enlighten us soon.
@@vladimirljubopytnov5193 Nah, just adjust all possible timings in the BIOS, Problem solved
I have MSI B450 Tomahawk MAX.
I had 8x4 DDR4 Kits and they were highly unstable even at 2133.
So now I have 16x2 (32GB) dual channel.
This is my message to everyone who owns MSI B450 Tomahawk MAX, get yourself dual channel memory and make your life easy.
Dude, never use 4 memories in a b450/b550/b650, those cards only have 2 efficient lanes, that means only 2 memories can run fast and the other two are standard. You have to go for an X570/X670, etc to use 4 clubs.
@@m1gu32011 Thank you! You saved me a lot of time!
@@m1gu32011 i have b550 with 4 sticks of ram 32gb in total should i just get 2x16gb instead of using 4 sticks? i also have xmp on set to 3200mhz
One does not simply "Call wendel" it more of a summoning really
Curious if there's a meaningful difference when dual rank is taken into consideration. ie: 4x8gb SR (as tested) vs 2x16GB DR. I personally have the latter in 3200CL14 and since you mentioned interleaving there may not be much of a performance difference between the above two configurations.
Yes dual rank should be the same as 4 sticks: th-cam.com/video/Q4aTB0k633Y/w-d-xo.html
Yes
Dual rank crucial ballistix with my 3700x outperforms single rank sticks, spent a lot of time researching and overclocking.
So they literally have to do benchmarks for the 5600x, 5800x, 5900x and 5950x with 3200,3400,3600,3800 & 4000 mhz ram with latencies varying from 14,15,16,17 to 18? All in difference configurations of 8x2, 8x4, 16x2 and 16x4... Wow they have to redo A LOT OF BENCHMARKS.
@@Traumatree Not if electrical interference from the empty slots reduces your overclock. 4 single-rank sticks on a dual-channel platform are effectively dual-rank. So, if you're going dual-rank, you should prefer to populate all slots.
I love how rigorous each of your tests are. I can always count on this channel to have a great analysis, even of these strange little details!
WHEW IM GLAD I WATCHED RIGHT TO THE END. I almost added two more 16gb sticks to my list of things to go along with the ryzen 5900x I plan on purchasing. My 32gb 2x16 3400mhz im using with my 3700x should be fine... Wow thought I was gonna end up with 64gb of ram to get the max performance out of the new CPU.
64gb is a lot. I would opt to get 32gb at higher clocks. I'm running an old single rank kits of trident z 3866s at 1.5v and at 3733 16 16 20 38. 16gb total. Make sure to time them right. Mine get 45.2ns reads pretty happy about it for kits I got 5 years ago. The upside with 64gb is if you do compute workloads though so if you can go for it.. they look good but I'm one of the unfortunate people out of work right now with no compensation.
Considering that RAM price is low enough to make 64GB RAM something not so extravagant, could you test or give your opinion on 2x32 vs 4x16 ?
Based on the mention of Wendell saying 2x16GB is the sweet spot, it seems we can't necessarily assume that filling 4 slots is always better than using 2 slots (for the same total capacity).
Also, would X570 vs B550 make a difference ?
I would love to get an answer to this. I am in the same boat to get a 64GB kit
@Matthew Yeah an extremely clickbait and intentionally misleading title, I expect better of GN. This 4 stick nonsense is going to be a myth that is now going to be spread around for the entirety of Ryzen 5000's existence.
@@gromitnwallace5281 here's the thing, though, think about all the people out there with 2x8. They can literally just buy another 2x8 of the same stuff and raise their performance like 5-10%. THAT is the takeaway imo
And honestly, like $100 for that amount of improvement on what is probably a $1000 rig is not even unreasonable from a price/performance standpoint
@@chapstickbomber Oh the information is definitely useful, but GN presented it in such a way that both the title and first 20+ minutes of the video are misleading to everyone, until you watch to the very end to see the caveat 'oh yeah this actually only applies to 8x4 vs 8x2'
Problematic Video.
Only those who actually watch the whole 27 minutes, will hear that " 2x16gb sticks performs best of all ".
Others who usually skip to the benchmarks, will now believe that as a general rule, using 4 sticks, no matter the capacity, boosts your performance.
Damn, now I'm happy that I bought a 2 kits of 2x8gb (by mistake), the 5000 series will be a worth while upgrade in the future (going from 2700x). Thanks Guys for the video ! Keep it up
I'm at work, so I can't watch the video with sound fully, but I've been researching more about the whole 2 sticks VS 4 with the 5900x since recently after stumbling on a video testing that and 4 sticks being only a bit better in certain instances.... And recently trying to build a PC for a friend and thinking of selling him my RAM to get 2 sticks for myself (would be discounted and he does video editing) and get those 2 sticks to have some RGB lighting.
..... And also reading more and more about how pretty much Threadripper boards can use quad channel and dual channels can have trouble with 4 sticks........ Yeah, it's not easy.
''Accidentally'' 👌🏻😂
Nothing gets past Steve and the GN crew!
"Higher is better, more consistent is"
Your TECH JESUS is evolving
Congratulations! Your TECH JESUS has evolved into YODA STEVE!
Has to be a side effect from not sleeping for a week. 😂
After seeing this, I added a matching kit of GSkill Ripjaws V 3200mhz to my rig and saw the Mersene/ Prime95 iterations drop from 30ms to 17-18ms with my Ryzen 5 3600x at stock 3.8ghz. This knocked 8-9 days off my workload. haha GG Steve xD
I'll do the same. Thanks
I would say : single rank dram vs dual rank.
2 dual rank sticks will give you same results as 4 single rank sticks.
Ordering another 2x8gb of G skill Trident Z royal 3600mhz right now for my 5600X. Planned on upgrading to 32gb eventually anyway so I might as well now. GamerNexus literally builds my PC lmao.
This could be myths or rumors but I've read that silicone differences can cause issues between ram and you should only use a set kit of ram instead of buying 2 seperate 8 gig sticks or something
return and buy a full kit of 4x8 kit they are tested to run together sticks from a different box can cause issues
@@dragonwithak Sounds like a myth to me because every single build I've ever done I always start out with 8 or 16gb and upgrade later down the line to another set. Never had any issues. No difference between me buying 4x8gb 3600mhz set of G skill royal or 2 2x8gb 3600mhz set of G skill royal. As long as timing, speeds, and brands are the same there shouldn't be any issue. But who knows just because I never had an issue doesn't mean some people could.
@@MoizCOUK Bought these a year ago lmao returning is outa the picture, I have never had an issue upgrading to another set ever in any of my previous builds. G skill has always been reliable for me in that sense. Regardless it's not a big deal at all. Worst comes to worse if I do run into an issue I'll just return the set I recently purchased and buy a set of 4. Easy peasy.
I thought I had RAM nailed for my next build (will be done within a month) . Gaming is only a little important, I want to run DaVinci Resolve for editing and rendering 2k vids and later 4k vids. The plan was to start with 2 sticks of 32 GB each and then add 2 more later if needed. Looks to me like single ranked is totally out of the question for me. I just have to get Ryzen optimized RAM wherever I can get it, apparently, and not worry about this. Please don't tell me I don't need 32 GB for projects, I fill 16 GB just going overboard with Firefox tabs.........like when looking at the right side of TH-cam to pick stuff for later viewing.......for PC research.
Steve i think you made a mistake at 23:57, in his video with AnandTech's Ian they say that 2 sticks of dual-rank memory work better than 2 sticks of single rank, while youre saying that 2x16 GB single-rank works the best while quoting Level1Techs.
Wendal's RAM may have been dual rank
@@Viewer19 yes in 2x16gb config you need it be DR 16gb stick like e-die ballistix 3600mhz and you can get rgb too! or no rgb in three colors.
it will then benefit from interleavin feature.
and then most mobo are daisy chain so 2x16gb is easier for memorycontroller to drive than 4x16gb or 4x8gb so this leads to higher clocks for memory with tighter timings.
Tech Jesus throws acronyms like they're used by everyone and their cat. Me who has no idea what they mean, I'm wondering why I'm still watching this. 😂
All I know is this guy really llikes shopping at TJ Maxx.
You Jay Linus Nick ... are the Go To Guys! My Next build before 2025 will be apart of what I learn from all of you!
You're essentially testing one vs two ranks per channel. This performance bump is expected (more or less). Just ask Buildzoid, he'll provide an explanation.
No, that is actually a good question. would 4x 1Rx16 be different from 2x 2Rx8 or 2x 2Rx16
@Valoro85 hey my dude try 1900 fclk but adjust vddg ccd to 1v and vddg iod to 1.15v and 1.2v soc it will likely help with stability.
2x16 in most sticks 4x8 as your 16GB stick will have 2x8GB one 8GB on each side. So essentially 2x16GB = 4x8GB
www.tomshardware.com/reviews/amd-ryzen-3000-best-memory-timings,6310-2.html
Already known in zen2
This is "Rank Interleaving" for you :P While one rank in a channel is busy refreshing or activating rows or such a second independent rank can be addressed by the IMC. The (lengthy) top comment on this somewhat explains this, but gets the channel width wrong: The Data-width on one memory channel is 64 Bit, so with 2 IMCs you get 128 bits of that, regardless of rank count. One would select a different rank via the CS-Bit ("Chip Select"), which is not a data- but a control-pin. Take a look at Micron's technical documents regarding their UDIMMs (and look for RDIMMs and LRDIMMs while you're at it!).
For weird Edge-cases you can even have multiple ranks on one Memory IC (usually that would be used in 3D-Stacked memory).
That top comment is confusing so many people with that "256 bit" thing...
Really interesting. Thanks for the info Steve and the rest of GN. Looking forward to watching a few more tests on memory config, but it's looking like 4 x CL16 at 3600MHz is the sweet spot again? Wonder how speed vs latency pans out this time too!
I feel like this is something AMD could have mentioned in a slide, they are certainly aware of it.
Anyone who knows how ram works is aware of it lol
@@davidfoote696 seeing as most people learn about pc's to build their own pc, now exactly how they work, I don't see how your reasoning makes any sense. people learn how ram benefits your pc, whether it be about it being the quantity of ram or speed.
I had just ordered a 8GBx4 kit the day before for my 5800X :)
I’ve had 2x16 3200c16 forever and just scored another 2x16 3200c16 to make a total of 4x16…this video makes me happy because it’s also the trident z rgb ram. Looks like I’m getting at least 5% gains with my 5900x at lower resolutions…I usually run 1080p, sometimes 1440p so yeah. I’m stoked for gains. I mainly just wanted 64gz for more multitasking, but this is a welcome addition. Great video, thanks for all the testing..👏🏼
how is it running? rule of thumb is usually to not just buy 2 sticks that didn't come off of the same manufacturing line, but to sell the 2 you have and buy 4 that did, so im just curious
@@JaxDagger I have the same configuration but I bought two packages separately G.Skill RIPJAWS V 3600mhz cl16,19,19,39 F43600C16D-32GVKC I put them at 3200Mhz, so far no blue screen and I play warzone 2.0 for hours with this seasonal heat and without any failure
It appears that Zen 3 is just capable of taking a significant performance advantage out of rank interleaving. All previous gen Ryzen and Core did not benefit from interleaving well. Watch Buildzoid's video on memory ranks for more detail.
dual rank ALWAYS had a per clock advantage. the main problem is that in many scenarios it came with a massive clock penalty that negated the per clock advantage.
@@ActuallyHardcoreOverclocking so do u still reckon 2x16 would still be better on 5000?
Can we have a 2 stick rank 2 vs 4 stick rank 1 testing?
I use 4*8 dimms on Taichi x370 with r5 2600x and I saw more stable performance in games even if I lost 4 ms. Latency from 64ns to 68ns as I running slower cos are mixed b-die hynix AFR
@@aditrex 4x8gb or 2x16gb yes
Really interesting find, thank you guys a lot! I am very curious to see if/how this influences the new AMD "Smart Access Memory" feature when the Radeon 6000 cards arrive for benchmarks.
Did you ever get around to doing a follow up to this? Found this really interesting. Also building a new pc and would like to know if 2x16 being the sweet spot produced the same results as 4x8. On another note just bought myself a Gamers Nexus mod mat and can’t wait to use it!
Just go 4x16 weeeeeeee
Hey man awesome channel, keep up the good work. Thought you might get a kick out of this.... Running my i5 3570k @ 5.1 for the last few years now, it's held up over the years, but every now and then needs a bump in voltage. The VERY FIRST WEEK I had it together, I had it gaming at a friends house, and something caught the CPU fan and blocked it from spinning, the chip sat throttling for about 3 hours above boiling point, great way to break in a new (at the time) chip.
Good find! I think you could have emphasized the difference between dual-rank and dual-channel though.
Thank you so much for clarifying that all memory can be ran dual or quad! Online stores always made it seem like they were special SKUs.
We'll need another review once RDNA2 comes out to see if having 4 sticks affect Smart Access Memory as well (Zen 3 + x570 + rx6000)
As a hardware engineer having worked several years on the development of PCIe controllers, I don't see any reason for that to affect the performance of SAM.
The Smart Access memory basically enables the CPU to see directly more GPU memory using a BAR reconfiguration that enables it to show more than 256MB through PCIe. Usually the PCIe BARs have to be that small because it's still 32b legacy during boot up. The gpu memory is then accessed through this bar window plus a memory bank register that is written whenever the upper bits of the address need to be modified. So if the memory can directly be accessed without needing to modify the upper address separately it may improve the access time, hence a slight improvement in performance when doing this. This has nothing to do with the cpu memory so changing the CPU memory configuration won't affect this.
As i recall with RAM manufacturers, the reaosn they sell sticks as a "quad channel kit" is because they are all tested together in 1 system to ensure that each stick is compatable with the settings being advertised. If you were to get 2 "dual channel" kits and put them in your system, there might be slight inconsistencies that makes them unstable to keep the same frequiencies and timings advertised, as they were not tested together as a 4.
Just osmehting to consider. Maybe a video to test this?
I only have a single data point, but I bought two 2x8gb kits for 4x8gb and they work fine.
Based on what I read, as long as the timings (the XMP profile) and the manufacturer (samsung/hynix/etc) are the same there's very little risk with mixing two 2x kits together.
And if you're buying both kits at the same time from the same place, it's pretty much a guarantee that they'll play nicely with each other.
This video has changed my life....I saw this video in morning, bought another 16GB stick and my games are getting improved fps. 144+ fps in doom eternal with RTX ON, previously I was struggling to keep 100 FPS. 100+ fps with DF optimized settings in cyber punk...god I missed so much frames in past 3 months. (I have 3060 Ti, i9 10850k@stock)
Dual-rank interleaving increases performance. 4 single-rank DIMMs run in dual-rank mode on a dual-channel CPU. You'll see the same improvement with 2 dual-rank 16 GB DIMMs on a dual-channel platform, but the remaining empty slots cause electrical interference and reduce your overclock. Similarly, you want to run 8 single-rank or 4 dual-rank DIMMs on a quad-channel platform.
Microelectronic shenanigans.
What a bunch of nonsense. I would say provide factual proof or don't talk. I'm talking about the electrical interference bs.
@@eazen Yeah, I also find that pretty wierd. You ususally dont get much interference from empty slots compared to actual components with electricity flowing through them being there
@@eazen special boards dedicated to RAM OC usually don't have 4 slots, only 2, same goes with mini-ITX boards that are usually better memory overclockers than ATX. 4 slots require longer traces as they are spaced out, as well as more complicated layout thus it affects signal integrity, but this occurs only at very high overclocks usually. Doesn't mean that empty slots are creating interference, so dude just used the wrong wording, but in general he is right
@@JackDisturber wrong wording. 🤣 Yes, very wrong wording. He was talking about something else entirely, and what you mentioned, that I'm aware of.
Hi Steve, I think you misspoke at the end and meant to say that 2x16GB dual rank dims are a little better than 4x8GB single rank dims. 2x16GB single rank dims (ex. Crucial Ballistic Max 2x16GB) are only 128-bit, so you are not getting the benefit of interleaving that way.
I would also be interested in seeing how 4x16GB dual rank dims perform in your future video because there is a good chance 512-bit would overload the memory controller and tank your achievable memory frequencies. I'm curious if the additional bandwidth of 512-bit at slower speed would be enough to overcome the FCLK penalty compared to 256-bit and higher speed. I suspect not...
TL;DW for everyone else...
For best performance you want to have a memory set up with 256-bit bandwidth meaning 2x dual rank dims(2x128-bit) or 4x single rank dims(4x64-bit). 2x single rank (ex. 2x8GB) can leave up to 10% performance on the table due to lack of interleaving.
The definition of a term "nerd" at 2:45, love it.
Thanks, this helps with my memory purchasing plans going forward for a 5600X when stock is replenished. Hopefully you will consider using the 8x2 and 8x4 configs with testing the 6800/6900 and RTX 3000. I appreciate and love all you do and share with us!
im buying the 5600x u bought 8x4? did u had problems with it ?
A really impressive video. Steve and GN showed today that how reliable their benchmarking methodology is. The attention to detail from you guys is off the charts. I have been a sub for 2-3 years. I have subbed for life. :)
Whenever I ask this question of 2 vs. 4 sticks, most of the time tech reviewers always tell me to go with 2 sticks for (2) reasons...(1). Running (4) sticks puts a much higher workload on the memory controller of the processor, and generates more heat. (2). You can buy (2) sticks with lower lower timing settings as compared to 4-stick kits, giving better performance. Now its looks like for the the Zen 3 that this isn't true, and 4-stick kits are better?
Thanks for this. It prompted me to download more ram but I have no idea if it’s improved things for me or not.
Speaking of memory timings, what happened to part 2?
I wish this "minor detail" was taken to account when comparing Zen3 to Alder Lake.
It's amazing that just using 4 sticks of RAM instead of 2 increases performance by up to 10%. That's a worthwhile upgrade that anyone can do!
It's rather obvious since it uses better the bandwidth of each chanel
Explanation is simple as that : 4x8Gb memory single rank double channel works as 2x16Gb dual channel double rank memory . And that means faster, because there are more 64 adresable units seen (double) and the memory controller can access the 64 units in parralel (for the dual rank memory).
And this is the same for AMD or Intel CPU's (not only Zen 3 ...)
I was getting a 32 with 16x2 but after someone reviewed my part selection for my Gaming/Streaming Single PC Build. They recommended I got with 32 with 8x4 because they said I would get increased performance if all 4 slots are filled rather than just two. I also increased my boot up storage from 500 GB to 1TB.
Good
Conclusion: Don´t compare different sites benchmarks, focus on the % difference between the products you want to compare for each reviewer. Also, 4x8 ram sticks has better performance than 2x8 but not because of memory capacity but because of number of ranks, so 2x16 dual rank might be even better. Great job and great finding, as always!
memory ranks*
@@dexter2392 You mean memory ranks instead of "memory channels"?
@@Inimigo85 yeah, 2x8 and 4x8 is still dual channel anyway
@@dexter2392 But I didnt say it wasnt dual channel :f
@@dexter2392 I understand what you mean now, I edited it!tks
For some reason, even though I know exactly what Steve is saying during these benchmarks, my brain just can't keep up with the actual words, but somehow I still understand everything.
It's like the words just bypass the conscious part of my brain.
It's a trip!
This sounds like a reason to get that 2nd pair of sticks past the significant other... GoodGuyAMD 😂
Get BZ in on this too. Can't wait for him to make a video of him rumbling on this for hours saying it's his nth attempt
Props to bz for his devotion to not editing though, even of he knows he's sending to GN that will edit. I imagine there's comedy gold in some fast talking Steve cuts... Maybe they're on patreon or something
Extremely informative as always. Wish the presenter would slow down a bit, your channel already does an excellent job of 'chaptering' the content to make it easy to find relevant information, there is simply no need to speak at 100 miles a minute. I could be in the minority, but I think I would rather have the video be 50-60 minutes instead of having to replay a segment 3, 4 or 5 times to follow what you're saying.
The probable reason is that even on same channel, with two memory module, memory controller can send a precharge commands to one stick, then execute memory transfer on the other, and by the time the transfer is finished, be able to do other commands on the first stick. Similarly, it probably utilizes better the burst transfer, by interleaving burst commands between two sticks. Very smart. It improves effective bandwidth and latency hiding.
I don't think it is Ryzen 5000 specific, and it is likely similar tricks are done on previous Zen CPUs. It should be rechecked. The longer load / store buffers on Zen 3, maybe make it more visible, but it should be similarly visible on previous Zen CPUs with multi threaded workloads that use memory a lot.
Haven't finished yet but I assume it's rank interleaving being especially beneficial to the latest architecture? So shouldn't dual rank sticks work just as well?
The way the cpu calls to memory has changed a large amount. So dual rank is a bit more "constrictive" on the cpu than 4 sticks
@@dralord1307 Interesting, on the other hand four sticks can have poor compatibility with certain memory layouts which will hurt signal integrity hence speed and timings. Either way having finished the video it seems Steve's on the right track, hope he talks to buildzoid as he and HUB both made videos about this behavior on zen2
@@teddygoboom1 if you really want to understand I suggest reading this,
www.anandtech.com/show/16214/amd-zen-3-ryzen-deep-dive-review-5950x-5900x-5800x-and-5700x-tested
@@teddygoboom1 4 sticks and stability was an issue all the way back in the Vishera/DDR3/FX CPU days. I'm not so sure dual ranked modules aren't also troublemakers as well.
@@dralord1307 it's the other way around, 2x 16 GB DR is saver and better than 4x8 GB SR. But the difference can be minimal.
I ran a Timespy benchmark with two sticks of G Skill Aegis 3200mhz, 8gb, 16-18-18-38, single-rank memory sticks installed. The Ryzen 5 5600x was at stock clock with increased boost capabilities through Asus PBO, and saw a peak of 4.85ghz during testing on all six cores at one time or another (no specific graphing available unfortunately). I then hopped on amazon and ordered two more identical sticks of memory. Once it arrived, I popped it in, checked that everything was working, and ran the test again with four ranks of memory. I saw a modest graphics score improvement but a whopping 8.3% spike in CPU score. This wasn't a perfect test. For one thing, windows updated between the time of the first score and when my ram came in the mail. Still, it's very interesting to me and I thought it would be worth sharing, regardless of the imperfections in the methodology. I may go ahead and pull the sticks and do a back-to-back retest on the same windows version to verify this result.
I’ve seen better performance going from dual rank to quad rank, i.e. 64gb on 3950x. Best of all, performance is much less sensitive to timings.
BTW: Judging performance based on Linpack, Geekbench, not games.
@@mrdwilkster Firstly, disclaimer: I’m no expert on RAM. My understanding is multi-rank memory has an advantage thanks to “rank interleaving”. Quoting Guru3D, “(Dual Rank) modules can often be a smidge faster thanks to a capability called “rank interleaving,” wherein the second memory rank can still perform work while the first is being refreshed for use.”
www.guru3d.com/news-story/amd-blogs-on-ryzen-memory-performance.html
Think of it as dual rank allows memory operations to occur in parallel or queue, like multithreading or hyper-threading is to CPUs. For memory this means there’s less penalty for high latency, versus single rank.
@@BurnsRubber any applications? Linpack and Geekbench are synthetics. Useful, but not in terms of what we should see elsewhere.
The 5900X and 5950X should get more from added ranks, though, as they should be able to queue up up to double the L3 misses if the 5600X and 5800X.
@@laurelsporter Yes benchmarking codes doesn’t always translate into real world gains and agree the 12 and 16-core should see more benefit as they can become memory bandwitch starved. Although synthetic, the calculations Linpack and Geekbench do are relevant to scientific and engineering software I use. Real world software isn’t easy to benchmark. Was mostly pointing out that switching from SR --> DR and DR -> QR seems to allow me to get away with looser timings without loss in performance, and can even add some performance if you can mostly maintain low latency timings.
How about 4 x 8 vs 2 x 16? will they performs similarly? I'm planing to go ITX and I hope that will not lose too much performance.
23:40
Wendel says 2x16gb sticks performs best of all !
just look at 2x8 vs 2x16, less than 1% difference
It will depend on if the 16 GB DIMMs are single rank (i.e. usually 8 DRAM chips total on the module) or dual rank (usually 16 DRAM chips).*
Single rank means less capacitance on the control and address lines allowing for higher speed whilst dual rank allows the memory controller to issue more memory operations in parallel by interleaving requests to the different ranks.
[*] I wrote usual as it depends on how many bits the individual DRAM chip width is as they can be 4, 8, or 16 per chip..
single rank 2 x 16 performs best according to level 1 tech
typically a quad channel kit will all be produced on the same prod line, exact same chip config for optimal compatibility in quad channel. You can get in sub optimal state with separate kits that may have not have exact same chips/chip configs. But yes as long as the ram sticks don't have too much variation, you can run four random sticks and you may still get quad channel stable.
Would love to see you have that conversation with Wendell on screen, I think that could be super informative for your viewers. Conversations between the most rigorous testing channel and The Janitor help me to think of questions I didn't even think to ask before seeing them. Thanks for all the great content and hard work!
So, when is the part two of memory overclocking coming out?