From what I read, x3d SKUs were going away as they were going to integrate it's benefits into the mainstream line. Hopefully they can address the heat and voltage issues with it so that overclocking can resume.
Zen6 will REDUCE latency going from one chiplet to the other, but VCache was previously stacked on CCD so latency to VCache will increase, not decrease. Also if L3 stays on die this means VCache will be turned into L4. This means L3 has to be searched and only then VCache is searched so youncan add another 10 ns. The advantages of X3D on IOD: a) The X3D hotspot issues are solved so Zen6 X3D can clock same as non X3D which gives around 10% higher clocks, but also higher power consumption. b) X3D is shared across all CCDs so we dont have CCDs like on the 7950X3D where VCache is only available on one CCD. c) iGPU can use VCache as well. So AMD can design more powerful APUs without using GDDR
@@Gindi4711 So, having just one chiplet which has vcache, could still be faster in games in some cases vs a 2 chiplet CPU which both have access to vcache?
@@nossy232323 As long as only half of the cores are used it does not matter if scheduler works correctly. For example 7950X3D is doing fine in current games, 7900X3D (2x6 cores) is apready loosing 10% gaming performance as only 6 cores are used.
Because the smaller zen4c cores provide most of the performance of a zen4 core it would seem logical to me to: 1 remove SMT because it is a security risk and the space can be used for a zen5c core 2 provide 1 Zen5 core per CCD because most of the time it is only 1 thread that is cpu limited 3 provide 15 Zen5c cores per CCD because secondary threads aren't at 100% cpu anyway and a Zen5c core is still faster than an SMT thread on Zen5 This keeps the design simple and you don't really need a thread director like Intel. Most tasks work just as fast on Zen5c as Zen5. And tasks that need a Zen5 core most often only need 1. When you go to 2 CCDs you get 2 high performance cores for the high end.
Good for AMD, but I'd rather they'd do this Big Leap for GPUs. In order to get a tangible uplift in performance and memory capacity over my RX570 8GB, I'd have to spend at least 300€ or 250€ for the same amount of memory that's been available on Polaris based cards for 7 years now. Going from my i7 3770 to a cheap AM4 system with R5 5600 would cost me a bit over 200€ and double or even triple my performance.
Gamers will almost certainly want to stick to x86 unless your absolute top priority is efficiency/battery life or something. But even that's not guaranteed to be a win for ARM + Windows. For performance, standard x86 stuff is what you'll want.
x86 isn't going anywhere for a long time. Not because ARM isn't up to it, but because the infrastructure for x86 is deeply rooted and will take a while for transitions to take place on a wider scale.
Not that I would suggest upgrading from a 5900X right now, but I wouldn't make purchase decisions based on RGT "leaks"... dude has been verrrry spotty.
they should pump out a prosumer mi300 - it is a natural thing to do for them and all the while they can continue to make successive generations of enterprise/ai focused apu - teaming up these rumored apu with pci-e v6 could help and be a nice adjunct to high end gpu - a monster apu for the masses would in effect make a new halo product especially when combined with gpu
For Ryzen(consumer)? If Zen 6 is not AM5, then yes. Otherwise I'd guess probably not. Not impossible, but I think it would require a short turnaround for at least one of these generations, but I wouldn't expect that. Seems we're dealing with more like 18-24 months per generation now.
AM5+ is a possible name, since they have a history of using the + moniker at times for new platforms that are still mostly the same old platform and are even backwards compatible to some degree. If they go that route, they'll make a lot of people pretty happy, since it will let them upgrade their motherboards first while they take their time getting the money together for the best CPU/GPU combo they can afford when they finally upgrade those as well. Otherwise, I suspect they'll stick to the current naming scheme which means it will probably be AM6 if it's a new socket entirely. But if the socket only needs a few extra pins added on, they might go the + route where the cpu's with those pins missing just auto-register as AM5 hardware, but the cpu's with the proper pin-out will register as AM5+. Oh, and since there is likely not going to be a new type of DDR memory coming out for a few years yet, DDR6 for example only, we probably won't see a new socket. I could be wrong of course, and probably am... but I suspect they'll not go the AM6 route until they need to implement whatever comes after DDR5. And... I hate to be the one to point out the coincidental correlation... but has anyone else noticed that the AM# tends to follow the DDR#? AM1... DDR1. AM2... DDR2. AM3... you get the point yet?
Crazy that no sources talks about an AI chip for Zen 5 or 6 yet... FFS AI is gonna be huge for Windows 12 in 2024... and Intel and AMD have nothing for AI in their next CPU?? What??
16 cores is pointless for gaming. I have a 5950x and I have disabled the second CCD. It means 3 cores get to 5.00GHz under PBO when only one did with both CCDs.
@@Safetytrousers but i have to work in multi thread environment, viturallization, compression tech as well. That is y interested in it. by the way game are in zip format it takes hours to unzip.
If AMD does a big/little but they also have an 8 core CCD, it's going to be more like 4 big and 12 little cores to make 16 because you can't have 8 big cores plus 8 little in the same die area as just 8 big cores...
Why? 4 big cores per CCD would suck in gaming forcing every game to use 2x CCD and everything have to go through the slow infinity fabric. It would make a lot more sense to have 8big on one CCD and the smaller ones on the second CCD. In not so demanding tasks in laptop one would use the smaller one for desktop and low power and latency. In games you'd use only one CCD and avoid the AMD dip. Anyways it'll take Microsoft YEARS to fix their thread director to handle this so I will not be buying this platform. I had enough with Intels big/small cores which also took years to fix.
I'm still not sold on the _"No high end discreet GPUs because we want to make cheap ones for everyone!"_ thing. I simply don't buy it. As for putting Zen 6 on AM6, whelp that would be a PR *DISASTER.* AM5 boards are already more than double what the previous 300-500 boards were for the most part (ei it was pretty easy to find a $60 b450m that still had 4 dimms, 3 NVMe slots, a x16 and x8 along with a decent amount of USB's and a solid VRM config. The only negative that stuck around the _just above ultra budget_ level boards was the downgraded audio. Now we have a *LOT* of people who paid $135, $150, $185 etc etc for the same feature set save getting 2 4.0 slots and 1 3.0 instead of only getting 1 4.0 and 2 3.0/SATA slots and the board is only good for 2 generations? No way. That is a surefire way to get people who probably already went ham on their power supplies for what they have to switch back to Intel because they are pissed. And rightfully so. AM5 was already way, way too expensive than it should have been and kept a ton of people on AM4 who otherwise would have upgraded. Then they come out with _budget_ 520's for $90 bucks? Come one. That's 50% more than a full featured board could be found on any given day in 2021. _shrug_ No shade at anyone for reporting something they heard, but unless the person telling me worked for or were the people making decisions, I wouldn't even waste the grain of salt on that one *UNLESS* they (AMD) has some type of refresh planned similar to the 2K series. If not, as I said, I have a very difficult time believing that they aren't releasing 5 and 6 on AM5
You can pick up an a620 motherboard for 80$ with 4 dimm slots and 3x m.2 slots. Secondly AM5 is more expensive because it's supporting faster connections which require more shielding, thicker copper and better connectors. If technical limitations prevent zen 6 from being the best it can be on AM5, I'd rather have AM6, it's not unlikely AM5 would see some continued support, just as AM4 is still receiving cpu releases to this day. Also higher end boards haven't changed that much in price, a good board in 2017 was 200+ and a good board on AM5 today is 200+ I paid 380$ for an x370e board in 2017, 380$ for a x570e board in 2021 and 480$ for a x670e board, we're not talking huge price differences here.
I honestly think that the big cores would be on the main substrate and the little cores can be stacked, or the little cores could be stacked on one another, or a cash die could be stacked. Because this would negate the benefits of AMD moving to a chiplet architecture. That or they are also using the tiling system with the die stacking.
@@Gindi4711 i don't think you've read what the guy wrote. you stack the small ones, the low power ones. it won't be enough power to cause heat issues. if you put a cooler on 5 stacked phone cpu's they'll still be fine, wouldn't they?? and you leave the big ones alone so they cool fine.
I would like to say the game Star Citizen is a huge step in the common future of gaming where abouts your RP and vast universe / meta of sorts is hand and hand of your gaming wants from a game and your money paid into your performance of the pc is actually respected. As of now the game is the only game on the market that holds its potential and value since its a updated game and not a buy it over and over again through a series like call of duty.
Moore's Law is Dead has also just been talking about this. th-cam.com/video/ebdg8ZIzOeE/w-d-xo.html Admittedly, he talks more about the Epic and server side, less about mobile and desktop. I wouldn't say he contradicts you, (he doesn't), but his information does seem to be a little different. LOL ~ By the bye, thanks for having some actual information in this clip, not just waffle and stream of consciousness... ~~ If I understand properly, Intel do Big cores and Little cores, and the little cores are like 1/6 the size of the big, and they run at a lower clock, they don't have hyper-threading, and there are some instructions they can't do. (AVX-512 instructions.) ~~ AMD do normal cores, and they do 'Dense' cores. Now a Dense core is about half the size of a normal core. It can do all the same instructions, it does have hyper-threading, it is more compact because the size of the caches (which has been growing generation over generation) has been cut back to levels more like zen2. I have conflicting info about the clock speeds but I hear they are quite power efficient. The people who really want these types of cores, are the server boys and the data centre people, who want racks of machines that feature 128 cores and 256 threads, 500 virtual-machines running on each processor. They do have possible application in mobile and low power cases, because they use much less silicon to achieve a similar result. If clocked lower, they do use very little power, but I hear mixed reports about whether they can be clocked up to 5 Ghz or so, or not. I also hear mixed reports about whether you can stick 3d cache onto them or not. If they're small because you stripped out 80% of your cache, then giving all that cache and then some back, using 3d, that would seem like a win. I hear that the Epic versions, are going to feature multiple I/O dies, with 2 CCXs per I/O. You could line 8 of these up in a row, with 16 CCXs and 128 or some large number of cores, but the real trick is to have a line of I/O, with a standard CCX on one side and a Compact cores CCX on the other, or a FPGA (hello Xilinix), or a new super AI inference module... Point is, these things are conceived to be modular and configurable. If AWS want one thing and Google want another and Meta want a 3rd, they can mix and match the modules to suit the application / customer. They have been designed with that concept in mind. As far as I know, the new Zen are all going to have a graphics system on every one of the I/O modules, but I'm not sure those are all going to be the same. As one example, it would help an APU to have a I/O die with 3d cache, which it used as V-cache... That 3d V-cache (Infinity Cache) has helped AMD / Radian's other graphics stuff plenty, so why not? I also hear from some, that the memory controllers are moving off the I/O and onto the CCX. Not everybody agrees on that point.... I would observe that does seem to contradict some of the 'theory' I've been told about previous generations, and using a slightly larger / older process for the I/O dies .... Much of the justification for that, had to do with off package comms and the memory controllers. If you can move the memory controllers back onto the CCX, then the only reason we had older process I/O dies was some business contract with Global Foundries. The technical expression for that is "Short memories." (Not the chips ~ us.)
Keep up the great work, Paul, and don't let the naysayers get to you. I've personally watched and followed you for years now and have seen you be extremely careful and accurate. Thanks for keep it fun and geeky interesting!! 🤓🤟🖖👍!!
I see there are CDNA and XDNA cards which are for AI, but could either of these run in a desktop along side an RDNA card? If not, it would be cool if AMD could get some eco system with SAM connecting a XDNA/CNDA with an RDNA GPU, ZEN CPU, and RAM together.
Why, when optimized properly RDNA can run AI or other tasks very well. Recently updates came out and a 7900xtx is faster than a 4080 for AI image generation. RDNA is faster than nvidia already for a lot of productivity stuff like autocad or video editing.
I’m conservative. I know AM5 will be for me. Just trying to wrap my mind around understanding the implications (reliability etc.) of small cores mixed with large is not appealing to me 😊
I fully intend to dig in with a 5600 (maybe 5800x3d if they start out-performing the 6 core substantially) and 3090 until 2025. I should stop watching these tech videos until then but now i"m addicted after the pandemic 😂
Just don't upgrade until the game you really want to get doesn't run to your liking. If the games you currently play run fine they will run fine in 3 years.
7900 XTX runs all games without issues just short of path tracing which Nvidia sucks at anyway. 0.01% of the Internet is convincing each other that AMD sucks because someone else thinks RT is important lmao maybe think for yourself once in a while.
@@christophermullins7163 I gave AMD a pass for not having great RT on RDNA 2 and RDNA 3, but RDNA 4 and 5 don't have any excuses for poor RT performance. Intel Arc runs circles around the comparable AMD and Nvidia GPUs in its 1st generation. Intel will be coming out with a GPU that will rival the RX 7900 series and faster too. AMD needs to get their act together ASAP this late in the RT game. It's not a feature that's going away.
@@AshtonCoolman "poor RT in RDNA 4"? You are basing your disappointment on a product that has not even been released. Also RDNA already has Ampere level RT and if RDNA4 adds 15-30% on top of that it's in no way "poor" as you put it. Also we have to remember the way RT is implemented in RDNA arch is by the number of raster cores. They are not separate RT cores that can be increased individually of raster cores. Also the price. As if Nvidia's current $500 cards are RT monsters? No they are not because the dies are small like RDNA4 will likely be. As for Intel. They are nowhere in comparison. Still fixing the huge backlog of games, still struggling to release day 1 drivers. Sales volumes are meaningless too. I would not place any bets on Intel releasing 7900 series competitor and if they will it will be in 2024. Remember that 7900 was released in 2022. Intel is a mess right now.
everything in the world updates 15 to 30% on every release...what kind of leak is that,,,,...Let me leak you zen 25 from zen 24...it will be 15 to 30% faster, I am just not sure about the naming scheme yet, thing's may change and games might change and I prefer to let you know when I know better. dude the whole deal about computer electronics today is if you can double or close to double performance or the efficiency from one gen to another.... 30% is a thing of the past so stop saying 15 to 30 % on every leak u promote on ur channel, that is not a leak , we dont need to watch your videos to expect 15 to 30 % performance increase in stuff.
So first, more "Intel Like" in the way MCM works means moving to tiles butted up against each other and direct connects. This is after all a joint patent between different companies INCLUDING AMD, or it's an open connection architecture so any company can use it. With direct connects there are WAY too many possibilities on how AMD could do Zen 6 desktop and this is why there is little information about it and I don't think AMD wants ANY of this to get leaked. AMD HAS to incorporate AI into the CPUs. Even Windows OS uses AI. What is it, Win11 23H2? Whatever the new release is Microsoft added a LOT of AI functionality into the OS. So the CPU should be the hardware running it or it will be slower if this has to bounce between a GPU and CPU. So that's one. But AMD has a problem with the IF and how fast it can go. It's already problematic for Zen 4 and why I won't buy that gen and because AMD is going to leave the IOD alone for Zen 5 I won't buy that either. It adds too much latency when data has to pass between cores on different chiplet and YES THIS IS GOING TO HAPPEN when you have multiple core chiplets and programmers can't fix that. The thread scheduler in the OS can minimize it as much as possible keeping all the threads of a certain app on one core chiplet, but you look at the 5950X and gaming and you see threads scheduled for cores on both chiplets, or you're rendering on the CPU, or running other tasks that can generate large numbers of threads such as compression/decompression and data has to move between the two chiplets so PLEASE don't give me blah blah blah programmer blah blah blah. No, thread scheduler, that's it, or for some apps you can limit the thread count and then the scheduler MIGHT keep all the threads on the same core, if you're single tasking, but for lots of users they're doing multiple things on their system at the same time in which case the scheduler probably won't keep the all threads for all apps on the same chiplet. But memory access is not good being at 6GT/s when DDR4 is supposed to get closer to 8GT/s. And even though the mem controller can handle faster kits, that doesn't make the IOD get the data to the core chiplets any faster since that has to pass through the IF, running at 3GHz. I think Intel tile MCM architecture is going to be much better and it's going to allow Intel to beat out AMD CPUs in many ways. There's so much AMD could do changing to possibly direct connects that it's too much to cover. But I don't see why the thin graphics would end up on the core chiplet. The core chiplet would be a better place for the memory controller.
Thread scheduler schedules games or active applications on the fastest clocked cores, all multi ccd ryzen chips have one ccd clocked higher than the other by a couple hundred mhz. The latency you mention is when a thread jumps from one ccd to the other, which isn't really a thing for multi core workloads like rendering or compress/decompress. If it was, the 7950x wouldn't slaughter the 13900k for compress/decompress, but it does.
@@PineyJustice We'll agree to disagree. Anytime you send data from point A to point B but that data has to go through a multiplexer first, you add latency. With the speed AMD is running Zen 4, compression/decompression should be faster than it is NOW with something like a 7950X and the faster the cores get, the more pronounced the latency will be. You're saying that core 0 has to pass data to core 10, and on the way that data leaves a core chiplet, goes into another chiplet, gets multiplexed, leaves that chiplet and goes into another chiplet and then on to point B suffers no loss in time. AND during that time that data is moving at most about 3GHz which is only about 50% of the speed data moves inside the core chiplets. Yeah we're going to agree to disagree. It' that same issue that limits memory speed. Data from DRAM, REGARDLESS of if a memory kit is 7200MT/s can only pass data through the infinity fabric at 6000MT/s or 3GHz since that's about as fast as you can reasonably clock the IF and data from memory has to go through the IF before it goes to a core chiplet. Also for various multi-threaded workloads the 14900K which is a 32 thread CPU just like the 7950X is close enough to the 7950X that I find it sad a CPU with 16 EFFICIENCY cores can be almost the same, actually beating it a lot, as an AMD CPU running 16 SMT POWER cores, and those efficiency cores are running slower, which helps to negate the advantage Intel has from those 16 e cores not being SMT (SMT rarely doubles the performance). Feel free to rewatch GN's review of the 14900K and look at Blender and Compression results. Or, go to Techpowerup's review of the 14900K and in multi-threaded workloads they show the 14900K beats the 7950X on a regular basis. So basically your memory of day 1 reviews is pretty shot. Maybe you're confused because a lot of power consumption diagnosis gets thrown in, but that's the difference in TSMC N5 vs. Intel 7 and how fast each company chooses to clock their CPUs. Also, you were completely wrong on saying the 7950X slaughters the 13900K in compression/decompression. Using WinRAR both the 13900K and 14900K are faster than the 7950X and using 7-Zip, it's only decompression where AMD wins and that's the difference in instruction logic, how each core performs each specific instruction. For the BEST testing results since they're about the most complete, look at Techpowerup. There you can see all the results without listening to the sarcastic commentary of a reviewer and get the REAL, unbiased picture of what's happening. When Arrow Lake comes out, if AMD isn't on a new MCM design they're going to get FLATTENED. 3 - 3.5GHz isn't fast enough for a clock speed for their infinity fabric which is a multiplexer on the IOD. If they can magically DOUBLE that speed on that IOD they'll help to negate its problems, but the real issue is as AMD develops ever faster cores, having an infinity fabric gets harder and harder to compete with against the likes of Intel who is going to beat the snot out of Zen 4 once they get on their Intel 4 process node for desktop. But you'll get a taste of it very soon, at least for laptops although Meteor Lake doesn't have all the changes Arrow Lake will. Mind you, I haven't bought an Intel CPU since the 1990s so it's not like I'm an Intel fanboy. I just take all the emotion out of all this and look at the DATA, something you have failed to remember obviously by your statements. Then add to this there's been at least a couple times now that Paul was talked about something AMD has said where their Infinity fabric is creating problems for them. Uh, yeah? They need to clock it at least 1.5X the speed it does now but that won't fix latency issues during core to core data transfers when cores are on separate chiplet. It will HELP, but it won't fix it.
@@johndoh5182 Normal workloads do not switch from CCD0 to CCD1. Threads don't interact like you think they do. The latency is specifically when a thread is transferred to another core, and even then the latency hit isn't that bad, it's about 4x higher than going from cache on the same CCD, but still far faster than the P to E core latency on intel CPUs which have stutter issues even in desktop applications. And phoronix is a better testing site, which does show the 7950x is handily ahead in many areas.
AMDCOREES ARE AN DHAVE ALWASY BEEN E CORS just as Intels since ZEN 1... there is no such thing as e cores for AMD . they might have hybrid architecture but what you call E cores will perform just as good as P cores....It is Intel P cores that are troglodyte antiques
Become? Anyways, you'll notice ALL these tech rumor channels run by nobodies all pander heavily to AMD hardcores/fanboys. Those sorts of people all seem to drink up hype and koolaid much more than others.
@maynardburger Very true. They all have a hivemind mentality in controlling a narrative that Intel is bad, and AMD is good when we know that both companies have done shady business practices and hosed consumers. But I just hate when AMD F's up they are never called out and are almost awarded for it and gets a free pass, but whenever Intel F's up, they are immediately called out and are bashed for it. There are definitely double standards that exist in the PC ecosystem when it comes to AMD vs. Intel and Nvidia vs. everybody
@@Serandi1987 thats when gta6 would come out on pc most likely we will definitely need gddr6 for gta6 im using 12600k and 7900xtx as of right now have you play gta5 is worth buying for 15 dollars on steam
@@Jason_Bover9000 yes I think DDR6 will be released next or maximum in two years. I didn't play GTA 5 but of course it would run absolutely great on yours setup! 👍🏻
If Zen 6 is not on AM5, that's disappointing.
True if that happens I will wait for am6 im using intel right now
Zen 7 should be am6 and and zen6 should be am5
I ain't buying nothing. This stuff is getting out of hand
@@oakfig true but if you want the best like to play gta6 there you go
If it's on am5 - 2 channels memory amd can't increase core count.
That's nothing - just wait until Zen 14.
Zen 6 is the last Zen... after that AMD changes the name.
@@lolmao500 to what
@@Jason_Bover9000To Ben.
@@Jason_Bover9000 "nen"
Will regular L3 be moved from CCD to IOD as well?
Or will this be a 4 level cache system with X3D being the L4?
From what I read, x3d SKUs were going away as they were going to integrate it's benefits into the mainstream line. Hopefully they can address the heat and voltage issues with it so that overclocking can resume.
Zen 6 is going to eliminate latency.
You can't eliminate latency. But, the design will certainly reduce it.
@@26Guenter Lol. That's what I meant. It would be one hell of a CPU that eliminates latency! Id camp out to buy that one!
Zen6 will REDUCE latency going from one chiplet to the other, but VCache was previously stacked on CCD so latency to VCache will increase, not decrease.
Also if L3 stays on die this means VCache will be turned into L4. This means L3 has to be searched and only then VCache is searched so youncan add another 10 ns.
The advantages of X3D on IOD:
a) The X3D hotspot issues are solved so Zen6 X3D can clock same as non X3D which gives around 10% higher clocks, but also higher power consumption.
b) X3D is shared across all CCDs so we dont have CCDs like on the 7950X3D where VCache is only available on one CCD.
c) iGPU can use VCache as well. So AMD can design more powerful APUs without using GDDR
@@Gindi4711 So, having just one chiplet which has vcache, could still be faster in games in some cases vs a 2 chiplet CPU which both have access to vcache?
@@nossy232323 As long as only half of the cores are used it does not matter if scheduler works correctly. For example 7950X3D is doing fine in current games, 7900X3D (2x6 cores) is apready loosing 10% gaming performance as only 6 cores are used.
Because the smaller zen4c cores provide most of the performance of a zen4 core it would seem logical to me to:
1 remove SMT because it is a security risk and the space can be used for a zen5c core
2 provide 1 Zen5 core per CCD because most of the time it is only 1 thread that is cpu limited
3 provide 15 Zen5c cores per CCD because secondary threads aren't at 100% cpu anyway and a Zen5c core is still faster than an SMT thread on Zen5
This keeps the design simple and you don't really need a thread director like Intel. Most tasks work just as fast on Zen5c as Zen5. And tasks that need a Zen5 core most often only need 1. When you go to 2 CCDs you get 2 high performance cores for the high end.
"I'm hearing" "I'm hearing" - no, you're just making guesses and passing it off like you're connected.
I’m happy they are going down an efficiency route, handheld gaming is going to just get better and better.
2:29. I thought this was about Zen6
16:20 ZEN 6 slides
I’m sticking with AM4 for a long time. It spanks everything I throw at it.
Good for AMD, but I'd rather they'd do this Big Leap for GPUs. In order to get a tangible uplift in performance and memory capacity over my RX570 8GB, I'd have to spend at least 300€ or 250€ for the same amount of memory that's been available on Polaris based cards for 7 years now.
Going from my i7 3770 to a cheap AM4 system with R5 5600 would cost me a bit over 200€ and double or even triple my performance.
6800 xt. 16gb vram and 20 times the Performance
What would be the implications to gaming or for Windows of ARM CPUs for AMD/Nvidia and when or will it happen?
Gamers will almost certainly want to stick to x86 unless your absolute top priority is efficiency/battery life or something. But even that's not guaranteed to be a win for ARM + Windows. For performance, standard x86 stuff is what you'll want.
x86 isn't going anywhere for a long time. Not because ARM isn't up to it, but because the infrastructure for x86 is deeply rooted and will take a while for transitions to take place on a wider scale.
All these big "changes" when in reality it's still just a x86 CPU consisting of two CCD and one IO die....
It would be awesome if AMD could release Zen 6 on AM5 but if we get to see 16 core CCDs before AM6, I'll be happy enough.
Sounds good, but I feel the bottleneck is developers needing to optimize better.
Thanks for the video! Since a new socket will probably be required I guess I'll wait to upgrade the 5900X.
Not that I would suggest upgrading from a 5900X right now, but I wouldn't make purchase decisions based on RGT "leaks"... dude has been verrrry spotty.
they should pump out a prosumer mi300 - it is a natural thing to do for them and all the while they can continue to make successive generations of enterprise/ai focused apu - teaming up these rumored apu with pci-e v6 could help and be a nice adjunct to high end gpu - a monster apu for the masses would in effect make a new halo product especially when combined with gpu
The socket could we end up seeing both zen 6 and 7 zen on same socket ????
For Ryzen(consumer)? If Zen 6 is not AM5, then yes. Otherwise I'd guess probably not. Not impossible, but I think it would require a short turnaround for at least one of these generations, but I wouldn't expect that. Seems we're dealing with more like 18-24 months per generation now.
AM5+ is a possible name, since they have a history of using the + moniker at times for new platforms that are still mostly the same old platform and are even backwards compatible to some degree. If they go that route, they'll make a lot of people pretty happy, since it will let them upgrade their motherboards first while they take their time getting the money together for the best CPU/GPU combo they can afford when they finally upgrade those as well. Otherwise, I suspect they'll stick to the current naming scheme which means it will probably be AM6 if it's a new socket entirely. But if the socket only needs a few extra pins added on, they might go the + route where the cpu's with those pins missing just auto-register as AM5 hardware, but the cpu's with the proper pin-out will register as AM5+.
Oh, and since there is likely not going to be a new type of DDR memory coming out for a few years yet, DDR6 for example only, we probably won't see a new socket. I could be wrong of course, and probably am... but I suspect they'll not go the AM6 route until they need to implement whatever comes after DDR5. And... I hate to be the one to point out the coincidental correlation... but has anyone else noticed that the AM# tends to follow the DDR#? AM1... DDR1. AM2... DDR2. AM3... you get the point yet?
Can someone explain to the n00b how this isn't arm?
CCD could be anything really and desktop will use any variant that makes sense.
Crazy that no sources talks about an AI chip for Zen 5 or 6 yet... FFS AI is gonna be huge for Windows 12 in 2024... and Intel and AMD have nothing for AI in their next CPU?? What??
It's already in the laptop chips, desktops are better off running it on the GPU than CPU.
Never trust windoze😂
Great vid as always. Cant wait to see 3Dvcache variants of future platforms.
when 5950X3D will be released
16 cores is pointless for gaming. I have a 5950x and I have disabled the second CCD. It means 3 cores get to 5.00GHz under PBO when only one did with both CCDs.
@@Safetytrousers but i have to work in multi thread environment, viturallization, compression tech as well. That is y interested in it. by the way game are in zip format it takes hours to unzip.
If AMD does a big/little but they also have an 8 core CCD, it's going to be more like 4 big and 12 little cores to make 16 because you can't have 8 big cores plus 8 little in the same die area as just 8 big cores...
What if, 8 big and 16 little??
Why? 4 big cores per CCD would suck in gaming forcing every game to use 2x CCD and everything have to go through the slow infinity fabric. It would make a lot more sense to have 8big on one CCD and the smaller ones on the second CCD. In not so demanding tasks in laptop one would use the smaller one for desktop and low power and latency. In games you'd use only one CCD and avoid the AMD dip. Anyways it'll take Microsoft YEARS to fix their thread director to handle this so I will not be buying this platform. I had enough with Intels big/small cores which also took years to fix.
Time to build a new PC baby
Wait untill DDR6
Why? He said there was not going to be a massive leap lol.
@@impuls60 this is a massive leap for what I have lol
@@impuls60Imo the real massive leap will be RDNA 5.
I'm still not sold on the _"No high end discreet GPUs because we want to make cheap ones for everyone!"_ thing. I simply don't buy it. As for putting Zen 6 on AM6, whelp that would be a PR *DISASTER.* AM5 boards are already more than double what the previous 300-500 boards were for the most part (ei it was pretty easy to find a $60 b450m that still had 4 dimms, 3 NVMe slots, a x16 and x8 along with a decent amount of USB's and a solid VRM config. The only negative that stuck around the _just above ultra budget_ level boards was the downgraded audio. Now we have a *LOT* of people who paid $135, $150, $185 etc etc for the same feature set save getting 2 4.0 slots and 1 3.0 instead of only getting 1 4.0 and 2 3.0/SATA slots and the board is only good for 2 generations? No way. That is a surefire way to get people who probably already went ham on their power supplies for what they have to switch back to Intel because they are pissed. And rightfully so. AM5 was already way, way too expensive than it should have been and kept a ton of people on AM4 who otherwise would have upgraded. Then they come out with _budget_ 520's for $90 bucks? Come one. That's 50% more than a full featured board could be found on any given day in 2021.
_shrug_ No shade at anyone for reporting something they heard, but unless the person telling me worked for or were the people making decisions, I wouldn't even waste the grain of salt on that one *UNLESS* they (AMD) has some type of refresh planned similar to the 2K series. If not, as I said, I have a very difficult time believing that they aren't releasing 5 and 6 on AM5
Both Polaris and rdna 1 lacked high end parts. It wouldn't be unprecedented
You can pick up an a620 motherboard for 80$ with 4 dimm slots and 3x m.2 slots. Secondly AM5 is more expensive because it's supporting faster connections which require more shielding, thicker copper and better connectors. If technical limitations prevent zen 6 from being the best it can be on AM5, I'd rather have AM6, it's not unlikely AM5 would see some continued support, just as AM4 is still receiving cpu releases to this day. Also higher end boards haven't changed that much in price, a good board in 2017 was 200+ and a good board on AM5 today is 200+ I paid 380$ for an x370e board in 2017, 380$ for a x570e board in 2021 and 480$ for a x670e board, we're not talking huge price differences here.
I honestly think that the big cores would be on the main substrate and the little cores can be stacked, or the little cores could be stacked on one another, or a cash die could be stacked. Because this would negate the benefits of AMD moving to a chiplet architecture. That or they are also using the tiling system with the die stacking.
AMD will not stack cores on other cores. This will create huge hotspot issues.
@@Gindi4711 i don't think you've read what the guy wrote. you stack the small ones, the low power ones. it won't be enough power to cause heat issues. if you put a cooler on 5 stacked phone cpu's they'll still be fine, wouldn't they?? and you leave the big ones alone so they cool fine.
I would like to say the game Star Citizen is a huge step in the common future of gaming where abouts your RP and vast universe / meta of sorts is hand and hand of your gaming wants from a game and your money paid into your performance of the pc is actually respected. As of now the game is the only game on the market that holds its potential and value since its a updated game and not a buy it over and over again through a series like call of duty.
It's funny that those ray tracing before and after hardly differ. No wonder no one cares about it yet.
Not buying a 96 bit GPU, and not spending a fortune on a 192bit GPU so it looks like this gen could be DOA for this guy..sucks.
Moore's Law is Dead has also just been talking about this.
th-cam.com/video/ebdg8ZIzOeE/w-d-xo.html
Admittedly, he talks more about the Epic and server side, less about mobile and desktop.
I wouldn't say he contradicts you, (he doesn't), but his information does seem to be a little different.
LOL ~ By the bye, thanks for having some actual information in this clip, not just waffle and stream of consciousness...
~~ If I understand properly, Intel do Big cores and Little cores, and the little cores are like 1/6 the size of the big, and they run at a lower clock, they don't have hyper-threading, and there are some instructions they can't do. (AVX-512 instructions.)
~~ AMD do normal cores, and they do 'Dense' cores. Now a Dense core is about half the size of a normal core. It can do all the same instructions, it does have hyper-threading, it is more compact because the size of the caches (which has been growing generation over generation) has been cut back to levels more like zen2. I have conflicting info about the clock speeds but I hear they are quite power efficient.
The people who really want these types of cores, are the server boys and the data centre people, who want racks of machines that feature 128 cores and 256 threads, 500 virtual-machines running on each processor. They do have possible application in mobile and low power cases, because they use much less silicon to achieve a similar result. If clocked lower, they do use very little power, but I hear mixed reports about whether they can be clocked up to 5 Ghz or so, or not. I also hear mixed reports about whether you can stick 3d cache onto them or not. If they're small because you stripped out 80% of your cache, then giving all that cache and then some back, using 3d, that would seem like a win.
I hear that the Epic versions, are going to feature multiple I/O dies, with 2 CCXs per I/O. You could line 8 of these up in a row, with 16 CCXs and 128 or some large number of cores, but the real trick is to have a line of I/O, with a standard CCX on one side and a Compact cores CCX on the other, or a FPGA (hello Xilinix), or a new super AI inference module... Point is, these things are conceived to be modular and configurable. If AWS want one thing and Google want another and Meta want a 3rd, they can mix and match the modules to suit the application / customer. They have been designed with that concept in mind.
As far as I know, the new Zen are all going to have a graphics system on every one of the I/O modules, but I'm not sure those are all going to be the same. As one example, it would help an APU to have a I/O die with 3d cache, which it used as V-cache... That 3d V-cache (Infinity Cache) has helped AMD / Radian's other graphics stuff plenty, so why not?
I also hear from some, that the memory controllers are moving off the I/O and onto the CCX. Not everybody agrees on that point.... I would observe that does seem to contradict some of the 'theory' I've been told about previous generations, and using a slightly larger / older process for the I/O dies .... Much of the justification for that, had to do with off package comms and the memory controllers. If you can move the memory controllers back onto the CCX, then the only reason we had older process I/O dies was some business contract with Global Foundries. The technical expression for that is "Short memories." (Not the chips ~ us.)
Keep up the great work, Paul, and don't let the naysayers get to you. I've personally watched and followed you for years now and have seen you be extremely careful and accurate. Thanks for keep it fun and geeky interesting!! 🤓🤟🖖👍!!
I see there are CDNA and XDNA cards which are for AI, but could either of these run in a desktop along side an RDNA card? If not, it would be cool if AMD could get some eco system with SAM connecting a XDNA/CNDA with an RDNA GPU, ZEN CPU, and RAM together.
Why, when optimized properly RDNA can run AI or other tasks very well. Recently updates came out and a 7900xtx is faster than a 4080 for AI image generation. RDNA is faster than nvidia already for a lot of productivity stuff like autocad or video editing.
V-cache on the IOD sounds more like Intel's eDRAM than an L3 cache.
Who the hell wants little cores? Noooooo! Lol
I have a good idea: to plug TPU cards to many PCI-E lanes for becoming an brainless computer to an AI computer.
I’m conservative. I know AM5 will be for me. Just trying to wrap my mind around understanding the implications (reliability etc.) of small cores mixed with large is not appealing to me 😊
Your titles are all the sane dude. Big leap every generation....smokes and mirrors...
Grizzly Adams talks Tech
Super excited for DD2, where they clean up all the bugs in original Donald Duck
I fully intend to dig in with a 5600 (maybe 5800x3d if they start out-performing the 6 core substantially) and 3090 until 2025. I should stop watching these tech videos until then but now i"m addicted after the pandemic 😂
Just don't upgrade until the game you really want to get doesn't run to your liking. If the games you currently play run fine they will run fine in 3 years.
It's sad to see that RDNA 4 will also suck at ray tracing. 30% faster at best isn't going to help AMD's GPU reputation.
7900 XTX runs all games without issues just short of path tracing which Nvidia sucks at anyway. 0.01% of the Internet is convincing each other that AMD sucks because someone else thinks RT is important lmao maybe think for yourself once in a while.
@@christophermullins7163 I gave AMD a pass for not having great RT on RDNA 2 and RDNA 3, but RDNA 4 and 5 don't have any excuses for poor RT performance. Intel Arc runs circles around the comparable AMD and Nvidia GPUs in its 1st generation. Intel will be coming out with a GPU that will rival the RX 7900 series and faster too. AMD needs to get their act together ASAP this late in the RT game. It's not a feature that's going away.
@@AshtonCoolman "poor RT in RDNA 4"? You are basing your disappointment on a product that has not even been released. Also RDNA already has Ampere level RT and if RDNA4 adds 15-30% on top of that it's in no way "poor" as you put it. Also we have to remember the way RT is implemented in RDNA arch is by the number of raster cores. They are not separate RT cores that can be increased individually of raster cores. Also the price. As if Nvidia's current $500 cards are RT monsters? No they are not because the dies are small like RDNA4 will likely be.
As for Intel. They are nowhere in comparison. Still fixing the huge backlog of games, still struggling to release day 1 drivers. Sales volumes are meaningless too. I would not place any bets on Intel releasing 7900 series competitor and if they will it will be in 2024. Remember that 7900 was released in 2022. Intel is a mess right now.
Who needs RT anyways?
i wait till Beast lake from Intel
Followed RGT for a few weeks now, granted I don't understand all the technical talk, but the content is always great. Thanks for the videos.
So basically AMD was lying when they say AM5 will be the platform for at least 3 generations (2026)
the Hyper CPUs
Nvidia is done 😂😂😂❤❤❤❤❤❤
everything in the world updates 15 to 30% on every release...what kind of leak is that,,,,...Let me leak you zen 25 from zen 24...it will be 15 to 30% faster, I am just not sure about the naming scheme yet, thing's may change and games might change and I prefer to let you know when I know better.
dude the whole deal about computer electronics today is if you can double or close to double performance or the efficiency from one gen to another.... 30% is a thing of the past so stop saying 15 to 30 % on every leak u promote on ur channel, that is not a leak , we dont need to watch your videos to expect 15 to 30 % performance increase in stuff.
So first, more "Intel Like" in the way MCM works means moving to tiles butted up against each other and direct connects. This is after all a joint patent between different companies INCLUDING AMD, or it's an open connection architecture so any company can use it.
With direct connects there are WAY too many possibilities on how AMD could do Zen 6 desktop and this is why there is little information about it and I don't think AMD wants ANY of this to get leaked.
AMD HAS to incorporate AI into the CPUs. Even Windows OS uses AI. What is it, Win11 23H2? Whatever the new release is Microsoft added a LOT of AI functionality into the OS. So the CPU should be the hardware running it or it will be slower if this has to bounce between a GPU and CPU. So that's one.
But AMD has a problem with the IF and how fast it can go. It's already problematic for Zen 4 and why I won't buy that gen and because AMD is going to leave the IOD alone for Zen 5 I won't buy that either. It adds too much latency when data has to pass between cores on different chiplet and YES THIS IS GOING TO HAPPEN when you have multiple core chiplets and programmers can't fix that. The thread scheduler in the OS can minimize it as much as possible keeping all the threads of a certain app on one core chiplet, but you look at the 5950X and gaming and you see threads scheduled for cores on both chiplets, or you're rendering on the CPU, or running other tasks that can generate large numbers of threads such as compression/decompression and data has to move between the two chiplets so PLEASE don't give me blah blah blah programmer blah blah blah. No, thread scheduler, that's it, or for some apps you can limit the thread count and then the scheduler MIGHT keep all the threads on the same core, if you're single tasking, but for lots of users they're doing multiple things on their system at the same time in which case the scheduler probably won't keep the all threads for all apps on the same chiplet. But memory access is not good being at 6GT/s when DDR4 is supposed to get closer to 8GT/s. And even though the mem controller can handle faster kits, that doesn't make the IOD get the data to the core chiplets any faster since that has to pass through the IF, running at 3GHz.
I think Intel tile MCM architecture is going to be much better and it's going to allow Intel to beat out AMD CPUs in many ways.
There's so much AMD could do changing to possibly direct connects that it's too much to cover. But I don't see why the thin graphics would end up on the core chiplet. The core chiplet would be a better place for the memory controller.
Thread scheduler schedules games or active applications on the fastest clocked cores, all multi ccd ryzen chips have one ccd clocked higher than the other by a couple hundred mhz. The latency you mention is when a thread jumps from one ccd to the other, which isn't really a thing for multi core workloads like rendering or compress/decompress. If it was, the 7950x wouldn't slaughter the 13900k for compress/decompress, but it does.
@@PineyJustice We'll agree to disagree. Anytime you send data from point A to point B but that data has to go through a multiplexer first, you add latency. With the speed AMD is running Zen 4, compression/decompression should be faster than it is NOW with something like a 7950X and the faster the cores get, the more pronounced the latency will be. You're saying that core 0 has to pass data to core 10, and on the way that data leaves a core chiplet, goes into another chiplet, gets multiplexed, leaves that chiplet and goes into another chiplet and then on to point B suffers no loss in time. AND during that time that data is moving at most about 3GHz which is only about 50% of the speed data moves inside the core chiplets.
Yeah we're going to agree to disagree. It' that same issue that limits memory speed. Data from DRAM, REGARDLESS of if a memory kit is 7200MT/s can only pass data through the infinity fabric at 6000MT/s or 3GHz since that's about as fast as you can reasonably clock the IF and data from memory has to go through the IF before it goes to a core chiplet.
Also for various multi-threaded workloads the 14900K which is a 32 thread CPU just like the 7950X is close enough to the 7950X that I find it sad a CPU with 16 EFFICIENCY cores can be almost the same, actually beating it a lot, as an AMD CPU running 16 SMT POWER cores, and those efficiency cores are running slower, which helps to negate the advantage Intel has from those 16 e cores not being SMT (SMT rarely doubles the performance). Feel free to rewatch GN's review of the 14900K and look at Blender and Compression results.
Or, go to Techpowerup's review of the 14900K and in multi-threaded workloads they show the 14900K beats the 7950X on a regular basis. So basically your memory of day 1 reviews is pretty shot. Maybe you're confused because a lot of power consumption diagnosis gets thrown in, but that's the difference in TSMC N5 vs. Intel 7 and how fast each company chooses to clock their CPUs.
Also, you were completely wrong on saying the 7950X slaughters the 13900K in compression/decompression.
Using WinRAR both the 13900K and 14900K are faster than the 7950X and using 7-Zip, it's only decompression where AMD wins and that's the difference in instruction logic, how each core performs each specific instruction.
For the BEST testing results since they're about the most complete, look at Techpowerup. There you can see all the results without listening to the sarcastic commentary of a reviewer and get the REAL, unbiased picture of what's happening.
When Arrow Lake comes out, if AMD isn't on a new MCM design they're going to get FLATTENED. 3 - 3.5GHz isn't fast enough for a clock speed for their infinity fabric which is a multiplexer on the IOD. If they can magically DOUBLE that speed on that IOD they'll help to negate its problems, but the real issue is as AMD develops ever faster cores, having an infinity fabric gets harder and harder to compete with against the likes of Intel who is going to beat the snot out of Zen 4 once they get on their Intel 4 process node for desktop. But you'll get a taste of it very soon, at least for laptops although Meteor Lake doesn't have all the changes Arrow Lake will.
Mind you, I haven't bought an Intel CPU since the 1990s so it's not like I'm an Intel fanboy. I just take all the emotion out of all this and look at the DATA, something you have failed to remember obviously by your statements.
Then add to this there's been at least a couple times now that Paul was talked about something AMD has said where their Infinity fabric is creating problems for them. Uh, yeah? They need to clock it at least 1.5X the speed it does now but that won't fix latency issues during core to core data transfers when cores are on separate chiplet. It will HELP, but it won't fix it.
@@johndoh5182 Normal workloads do not switch from CCD0 to CCD1. Threads don't interact like you think they do. The latency is specifically when a thread is transferred to another core, and even then the latency hit isn't that bad, it's about 4x higher than going from cache on the same CCD, but still far faster than the P to E core latency on intel CPUs which have stutter issues even in desktop applications. And phoronix is a better testing site, which does show the 7950x is handily ahead in many areas.
Zen 6 will be incredible. We could see up to 5% IPC gains.
Your wallet will zen be #enlightened
AMDCOREES ARE AN DHAVE ALWASY BEEN E CORS just as Intels since ZEN 1... there is no such thing as e cores for AMD . they might have hybrid architecture but what you call E cores will perform just as good as P cores....It is Intel P cores that are troglodyte antiques
❤❤
RedGamingTech has defiently become a click baiting AMD fanboy channel.
RGT: ZEN 5, ZEN 5, ZEN 5 is gonna be massive
Nah my guy. Its just Intel does not had much of anything interesting except boosting its CPU to high heavens.
Went from Intel to AMD, trust me pal it is well worth it. However stick to Nvidia cards as for now.
Become? Anyways, you'll notice ALL these tech rumor channels run by nobodies all pander heavily to AMD hardcores/fanboys. Those sorts of people all seem to drink up hype and koolaid much more than others.
@@damienjeremytrotman84 This is just nonsense. Intel probably has one of the most interesting roadmaps in terms of technology over the next few years.
@maynardburger
Very true. They all have a hivemind mentality in controlling a narrative that Intel is bad, and AMD is good when we know that both companies have done shady business practices and hosed consumers. But I just hate when AMD F's up they are never called out and are almost awarded for it and gets a free pass, but whenever Intel F's up, they are immediately called out and are bashed for it.
There are definitely double standards that exist in the PC ecosystem when it comes to AMD vs. Intel and Nvidia vs. everybody
mb
5
First🎉
How
Your always first on his videos
@@Jason_Bover9000 I think it has to do with who gets a notifications first, I’ve also been subscribed since 2017. Lol
@@NovaDoll o ok when will am6 platform come out zen 6 or zen 7 im thinking zen 7
No one gives a shit 🎉
Will AM6 run with DDR6? Then I'll skip AM5 with my AM4👍🏻
It should
@@Jason_Bover9000 Hope so :D
@@Serandi1987 thats when gta6 would come out on pc most likely we will definitely need gddr6 for gta6 im using 12600k and 7900xtx as of right now have you play gta5 is worth buying for 15 dollars on steam
@@Jason_Bover9000 yes I think DDR6 will be released next or maximum in two years.
I didn't play GTA 5 but of course it would run absolutely great on yours setup! 👍🏻
@@Serandi1987 ok thanks man
Keep your hand still.