2990WX Threadripper Performance Regression FIXED on Windows*
ฝัง
- เผยแพร่เมื่อ 19 ก.ย. 2024
- *At least in cases like this one :D
[ sry, clickbait works, but see these awesome articles below v]
Full article, including the mentioned numa/"ideal cpu" dumps
level1techs.co...
Phoronix Windows & Linux comaprison
www.phoronix.c...
Ian Cutress' article on The Core0 mystery
www.anandtech....
Our Last Video on this:
• Threadripper 2990WX Pe...
**********************************
Thanks for watching our videos! If you want more, check us out online at the following places:
+ Website: level1techs.com/
+ Forums: forum.level1tec...
+ Store: store.level1tec...
+ Patreon: / level1
+ L1 Twitter: / level1techs
+ L1 Facebook: / level1techs
+ L1/PGP Streaming: / teampgp
+ Wendell Twitter: / tekwendell
+ Ryan Twitter: / pgpryan
+ Krista Twitter: / kreestuh
+ Business Inquiries/Brand Integrations: Queries@level1techs.com
IMPORTANT Any email lacking “level1techs.com” should be ignored and immediately reported to Queries@level1techs.com.
-------------------------------------------------------------------------------------------------------------
Intro and Outro Music By: Kevin MacLeod (incompetech.com)
Licensed under Creative Commons: By Attribution 3.0 License
creativecommons...
The W in 2990WX and 2970WX stands for "Wendell"
Like you say some of it is memory, but on Windows a lot of it is... Windows :D
Someone should look into the Pascal SLi + G-sync performance issues :). why does G-sync cause a large performance regression with SLi pascal cards www.extremetech.com/extreme/279259-nvidia-gpu-performance-craters-when-g-sync-sli-are-used-together
Oh holy hell, steve?! No kidding , it's encouraging to see you over here. Been following you awhile now , keep up the great benchmark work . Please cover this on your channel . More voices are needed for this issue.
Lol I was about to link Steve to this video before I looked down and saw that he'd beaten me to it
could you please redo the benchmarks for 2990wx
hope it gets fixed before my 2990WX arrives
Classic Windows 🤔
@Dominic Cerisano that helpfully saves space on your hard drive for you :D.
@Dominic Cerisano not installing linux saves even more space
@Dominic Cerisano as if amd drivers were good on linux, hell most drivers suck on linux
windows is trash, always has been
Ian cutress is a truly amazing journalist he is one of the few journalists that I actually read and follow closely.
"JUST carefully consider the price and performance, and BUY IT if you are in need of such item ." - Ian Cutress
I really hope that you submit this information to Microsoft and AMD, and perhaps some media outlets so it can get some publicity. That way the issue gets fixed in a timely manner. Great work you're doing here.
AMD should include Process Lasso (maybe AMD branded/verified) for the complex CPU packages. (or similar functions in their driversetup)
They did a deal for a free ramdriveprogram with Dataram when you bought Radeon Memory, and sometimes they give/include games with GPU`s, so why not include ProcessLasso and fully mitigate MicroSoft alltogether. (20 bucks is hardly that much for a Threadripper buyer, certainly not if it guarantees proper core utilisation)
It`s already half a year since the 2990WX has been released, before MS is going to include it in a patch it is going to take at least another 3 months because of "testing"/delaying.
So if AMD/you want to wait for MS to get of their ass you`ll be waiting about a year before you can use your whole CPU properly.
But we are currently in an economy of "don`t care", AMD didn`t care, MS OS-team and Office-team didn`t care, MB manufacturers didn`t care (they do automated memory timings and CPU clocking for example, why not scheduling too), Developer-software didn`t care (no mention from Adobe, enzzz...), Game software didn`t care.
So many industrial partners (nowadays acting like bankers) could have done this in 1-2 weeks , but didn`t. (so don`t expect them to suddenly change their ways)
{When (something wrong) Then (user error),
When (something wrong) AND (many users) Then (other developer error),
When (something wrong) AND (all users), Then (look into it) AND (don`t generate expectations to userbase) }
He will get further informing AMD. Microsoft is probably doing this intentionally or doesnt know and doesn't care. By intentionally I dont mean to intentionally reduce performance. I meam that possible but I just mean that Microsoft and Windows is always doing "funny stuff..." Well it is also possible they were problem solving possibly for intel CPU's with this behavior or exploiting them LOL!
@@TH-camDoxedMyRealName in b4 "was caused by fix in kernel for bulldoser scheduling"
@@seculi7757 well AMD should care because it's crucial to their success.
@@Megalomaniakaal Interesting I havd one of those. Ill have to mess around with it to see if it really works. Thx for the info!
When your scheduler is using more CPU utilization than the actual work scheduling.
+
Sounds like my job
This actually explains a lot, and fits the observations perfectly. Why Doesn't it surprise me at all that Windows is shuffling stuff around so much that nothing gets done.
it still count as making progress
xkcd.com/1906/
🙌
AMD should pay you or send you new cpu's to test AMD should be very happy that you found this :)
or just hire him
@Marc Jackson "MS needed a huge rewrite to fix properly" AFAIK it's not fixed in Windows as of right now....
MS should put back more knowledgeable people on their Windows team . Most of them are gone, a lot of them went to Azure. Like for example "Mark Russinovich"
@@mumblic OMG I have used so many of Sysinternals' tools, Mark Russinovich is an expert lever programmer for sure
I don't feel like we can or should blame anyone in the media (not that Wendell was). It would have been AWESOME if AMD tested this before they sent it out and didn't just fill the review guide with rendering benchmarks that worked well on Windows. They had to know this was largely a Windows issue and they could have easily shown that EPYC suffers the same performance issues. Would have made the performance at launch much easier for everyone to swallow.
All that said, great video. Wendell and white boards, I'm complete :D
The blame lies with the uneducated media with hardly any knowledge of the underlying architecture, simply running "gaming" benchmarks just to get clicks and views.
You would have to be the Geeky one of the two ... Tim? ;)
Dont think AMD has that much money they just needed to get this to market and deal with the small hitches later.
@Hardware Unboxed you guys better report on this to spread Wendells incredible work and let the news run through the wider tech community.
@@charlestwoo ☝ This.
This video should make tech headlines. I'm calling it!
Congratulations to you Wendell, and to everyone else involved. You did a fantastic job. I'm sure it took a lot of hours and effort, but it's fantastic that the community is actively helping to find and properly explain this bugs.
OMFG, just tested Process Lasso on my 5820K.
Before, when i had my CPU pegged at 100% pc was unresponsive and could barely move my mouse over the screen to reach task manager and see wtf was happening.
Now with PL in the background it behaves like before for one or two seconds and then like nothing is happnening.
Amazing.
I'll give Process Lasso a spin as well... to alleviate the dying pains of my ancient system before it becomes unusable for good.
Hopefully it'll last until my friend finally figures out how to get Looking Glass working on an eGPU with Razer Blades. Razer Blades. .... Razer Blades. From that point onward it's Fedora with no looking back. I will switch eventually, I swear.
I've been using it for years and it makes for a good multitasking machine. It does many things, and people who don't understand it think "bah, I can set affinity myself".. Sure. And have to do it every time you run anything. There are persistent priority settings not just for CPU, but IO, memory and it just goes on and on. You can just run it without touching any settings and see a decent increase. I play games without it and notice a lower number of cores using both of their threads (4 out of 8c, 8t out of 16t) and with ProcessLasso running, you can see it spread it over all 8 cores with all of their secondary threads barely running. It really is a very smart program. Well worth buying, and I did. I already owned ParkControl, but PL has this functionality (for the most part) built in. I would recommend anyone with a multi core CPU (likely everyone at this point).. try this.
@@larrygall5831 Are you suggesting that one can actually see an increase in, say, gaming FPS(for a non-niche example) on a modern 6 or 8 core CPU in cases where a game/application(especially older stuff) is programmed in a fashion that makes use of no more than 4 cores(if not only 2 even)? In this case, I'm assuming that PL is taking the workload from the application and redistributing it across all available cores in the CPU instead of only using the 2 or 4 cores the app would've only used if PL weren't superseding the original programming.
Or, is this more along the lines of offering relatively minor differences in performance in applications like this, but maximizing the efficient use of your system hardware(specifically the CPU) by spreading the work load across all available cores instead of, again, software only using 2 or 4 cores?
Forgive me for my lack of research about this at the moment, I'm just now discovering this. Either way, it sounds absolutely fascinating, and like something that may become VERY useful with my R5 1600 and, more importantly, the upcoming 8, 12, and 16 core mainstream Ryzen 3000 lineup(can't wait for CES!) and the up to 64 core TR 3000 lineup
@@Traumatree Ahhh, I see. Interesting. I'm definitely going to have to play with this.
@@RyTrapp0 That might be the case.
I've seen a certain game having the affinity issue. that is, performance improves if you manually change affinity when its running, doesn't matter if you just change, then change it back to how it was, it just works better. launching steam with a bat file that makes the games inherit the same affinities, does not seem to have any effect.
It is certainly something to look into, even though it might be unrelated to this issue.
I actually said at TR2 launch that the Windows scheduler wasn't tagging the cores properly within their ccx groups.
When the thread is moved over to a package that has no prior knowledge, the pipe is flushed and what looks like core utilisation is work getting restarted at penalty.
AMD always gets a black eye for innovation :c awesome work chaps.
Intel has gotten a fair few as well. Netburst (pentium 4) and itanium spring to mind. Itanium was going to be the 64-bit architecture; that’s why they didn’t bother making x64 CPUs themselves and let AMD go first. The bet they made with itanium was HUGE and it failed spectacularly.
I swear this was not 360p before I went live. Stupid youtube. :(
GamersNexus has had the same problem lately
Guffaw.
its standard wait hour or 2
Yeah the resolution will go up after a little bit of time
360p club
Two points:
1. I have a bridge to sell you if you think MS didn't know about this
2. Maybe building a 'sniffer' of sorts to monitor the threads continually to gather data and see any emergent patterns would be more beneficial at this point rather than trying to reset. I would imagine some significant percentage of these are being created and ended very rapidly.
Solid work on this though; seriously, really impressive.
Can't wait for the microsoft patch to fix this that ends up deleting all my files
omg, almost fell outa my chair laughing. (edit) some alcohol was in the mix too :D
Same here :) writing this from my Ubuntu desktop while drinking beer ( KDE Neon to be precise)
@@Ruhjuh Meanwhile im laughing from my chair running Windows as the last time I updated Ubuntu it bricked itself and files went missing.
@@silvy7394 It was probably your fault with configuration. When you are installing Linux you definitely should do two partitions to separate mounting points "/" and "/home/". That way if something goes wrong with your Linux installation just reinstall and mount "/home/" at your previous partition so you won't lose anything.
Clearly far superior solution than Windows and it's "Users" and "ProgramData" folders located on "C:\"
@@idontfeelsogood2063 So you're saying Linux cant properly install itself now?
Wow. Linux just got a whole lot more pathetic in my book.
*Just kidding but kind of not.* I know you're the average fanboy who plays the "its your fault" every time Linux bricks itself, which is all the time.
AMD: We got 32 Cores.
Windows: Never gonna work.
Wendel: Hold my beer.
Dude, amazing work. This all need to be taken to mind by both Microsoft and AMD. There is no longer any excuse therefore, for AMD's core management to not work well with windows. Impressive work, rock on dude.
*It's not a "bug" it's a "feature"...*
How else will Intel be able to compete with Amd.
Also Windows is using 50 percent of your CPU to process other "projects" when you expect it to be pegged at 100 percent LOL! Two birds with one stone!!! Set affinity in software any idiot can do that they wont notice or they will just think AMD sucks and start a hype train for us... Oh crap he looked at his performance and noticed their was a problem and set the affinity in task manager! Also he knows how to use the task manager! Power user detected suspend operations on this CPU for this important task LOL! We'll just wait and find a task he doesn't care to monitor so closely TROLOLOL!
And won't intel xeons or just any high core intel server socket CPU also win performance if they fix this? Or is it just with AMD?
Windows scheduler hints, written by Intel.
if (PROCESSOR_NAME != "Xeon") MAX_CORES = 16; else MAX_CORES = 4096;
Can you please try to benchmark linux running in hyper-v under windows host and see if results will be different ? Hyper-v is a type 1 hypervisor so if there is a performance bug in the kernel you might bypass the bug and prove your theory with higher confidence, and maybe localize the issue.
This would be an interesting test. I would expect the performance problem to remain. I really think of Hyper-V Server as a pseudo type-1 hypervisor. The fact that changing certain settings like the time launches traditional control panels and windows like you would see in Windows Server has always made me a little suspicious that it is just an even more stripped down version of Windows Server Core.
If only microshaft had an extra 20,000 employees to fix this issue.
Rofl , so unbelievably true. And that's what's so horrendously laughable.
Its ok they have the AI!
I'm not sure what I'm more impressed: the fact that you pretty much solved this puzzle, or the fact that I have seen incomplete answers about this problem on a local aluminum welding forum (they were just blaming Windows scheduler, but no mention about a bug).
Figuring this stuff out, especially on the spaghetti that is Windows, is seriously impressive. You might have even beat AMD to the punch.
I actually wouldn't be surprised.
@@texasdeeslinglead2401 They're a hardware company, the OS doing basic core scheduling is a Linux / Windows thing. Linux got it right off the bat, Windows is still struggling 2 years into Zen. Even when it's working "correctly", it's shuffling threads all over the place, very quickly and quite constantly. I see it daily.
@@larrygall5831 explain
I remember when Windows 95 didn't boot on some AMD CPUs because they executed the LOOP instruction too fast
I had the same problem.
Omg is that truth?!
@@Ag89q43G0HyA yes
WE NEED MORE OF THOSE VIDEOS!
Awesome video Wendall, I used the Bitsum app on My 1950x Threadrippper and saw a huge performance jump ... The work and research ya did for the new Threadripper is leading edge. I commend you for your hard work and effort to solve the hidden problem... Im a big fan of never leaving a stone un turned...when it comes to troubleshooting..
Windows = R.I.P for workload. good video. Linux Master Race.
Install Gentoo
Sure when it become popular, until then Linux is useless for common people.
@@Sturmbutcher Is it "masterrace" if it's common?
@@jordanwaughtal7649 No, it then becomes 'master racism'
@@jordanwaughtal7649 why does that matter? If it's superior to the competition, isn't it the master race?
Totally digging that L1 background .. and the schooling Wendell is dishing out.
Win10 + vray 2990wx, performance increase by around 25%, awesome work ! Thanks !
YES! This is why I'm a subscriber. Fantastic work from all involved in this investigation! Microsoft, get your sh*t together. If Linux can do it, you should be able to as well. I will try to spread this information and video as far and wide as possible. This needs the attention it deserves.
This is unbelievable investigative work. Well done to all involved! Hopefully it will be fixed now thanks to your efforts and this will make a massive difference to AMD's prospects.
I feel like anyone who down-voted this video must be either:
A) Computer illiterate
B) An Intel fanboy who hates AMD and wants them to fail.
Great video, what a fun puzzle.
If the issue gets solved by this, I really wish you and the others involved would get a reward for this even if the reward would be a small one.
I dont know jack about code and how most hardware actually works, but your explanayion of all this was so good i came to the same conclusion you did 30 seconds before you said it in the video. The first thing thst popped into my mind was that the kernel and its optimizations are designed with a single die per socket in mind. Then you said it and i felt smart, quality content
so much nerdiness here, i love it.
btw who cares about 360p, i enjoyed it anyway
this was something that im curious about too, thanks for being the one who investigates about this to go about in more detail!
M$ will "fix" it right when Intel launches chiplet CPUs.
I was thinking the same thing. I'll be upgrading my 1800x to a "3800x / 3850x" or whatever the new AM4 flagship is.. and passing on Intel again. I'm tired of everyone getting lazy and focusing on one CPU. AMD's market share will rise a bit more after this next gen 3 months from now, so it's getting to be less of a problem as time goes on.
Just that the same does actually exist with intel machines too............
maybe you should stop wrapping your whole head in tinfoil.
@@ABaumstumpf does it? intel doesn't have anything like the Threadripper/Epic design. as wendell himself said, this is the first time we've had designs like this with more then 2 numa nodes inside the same socket.
Not saying ms did this on purpose to help intel, but your claim that intel has it as well is false.
@@TheCountess666 Uhm - did i ever say that we had NUMA on a single socket before? No, but actually we already had - Intel Core 2 Quad and that too had those issues.
@@ABaumstumpf core 2 quad (and the pentium D that came before it as well) was just 2 sockets in a single socket. it functioned EXACTLY like a dual socket system. There was ZERO direct die to die communication, everything went through the northbridge.
AMD was also first by 2 years with the 64bit technology and Windows xp 64 was build based on that architect and that's why the folder in Windows set-up in all windows and drivers is called AMD64. Microsoft got tired waiting for Intel back then and went with AMD to make windows 64bit. AMD always innovates. I hope Microsoft doesn't want to loose more server % and actually do fix their shit fast.
There was "64-bit technology" long before AMD64 (since the 1960s, in fact). AMD was just the first to add 64-bit extensions to x86 (whereas Intel was trying to segment the market between 32-bit x86 and 64-bit IA64/Itanium)..
Every time I hear about a problem in Windows, I think i should learn how to use Linux.
It is pretty easy, just pick a distribution and run with it. Ubuntu/Solus for games (check the level1 linux for some info on configuring that) For development I prefer Manjaro since I can quickly and easily grab the latest libraries, Fedora or CentOS are probably the most corporate distros but at the end of the day just about any distro can be used like any other if you fiddle hard enough, just go for it. It is a very fun and rewarding ride :)
Linux is easy to use for people who just want to browse the web.
For tech savvy people it's a bit tricky to understand things "under the hood" which are very different from WinNT. But the learning curve is straight.
And for Gamers it is only usable with very strong hardware as the GPU drivers just don't work that well. (Not an OS issue, but still an issue for us customers). Also many games just won't work. Virtualization you say? BattlEye says "no"! Thank them for that. Apparently VMs can be exploited for cheating. Very sad they have to block VMware.
Eh, I was never able to get Netflix to run on Linux so I gave up.
@@n.i.5381 it is not a problem anymore, though I know well the Silverlight debachel you speak of. I know chrome just works, chromium, and Firefox probably just work as well now, though since I dropped my Netflix sub a few years ago I can't really test.
@@ithraldharzul6887 Maybe I'll revisit then as we dropped our cable co and use Netflix only on laptop.
Thank you for this. There were many times where I thought I was going to get lost, but you have a way of explaining everything perfectly. Than you.
Damn dude, PC Master Race indeed.
I watched the video and loved it, but had a hard time understanding all of it, due to lack of training. Maybe even ability.
I have to respect your work ethic here, you may have uncovered something kinda big.
I have no idea who you are or your history, and I usually have a hard time following you, but am glad I stumbled your channel.
Now I want to see where others take this, and I want to hear what Microsoft has to say about it.
Keep pushing the envelope.
And glad you did not do the typical AMD bad Intel good schtick.
AMD, give this nerd a medal.
Great work! Wonderful video and great explanation loved the white board and the breakdown keep it up!
This is the team up we needed. Wendell from Level1Techs and Ian Cutress from AnandTech
Legend. You are the best tech channel on YT. Empirical evidence via randomized experimental design. Chapeau!
I have learned so much in this short video. You gained a sub here. A lot went over my head but I am sure your video's are going to be very instructive and enlightening.
Intel + Microsoft Vs AMD ?
Holy unbelievable crap, thank you. I thought I'd left my sanity behind somewhere trying to figure out where my performance losses were coming from.
That's because Microsoft sits on three face of Intel. Both don't innovate anymore. It took AMD to get Microsoft to move to 64 bit. Otherwise they were waiting for Intel.
LOL Windows XP 64 bit sucked! It would eventually eat all system resources and run slow as crap! The solution was to reinstall the OS not sure if they ever fixed that I had to revert to 32 bit after the third install. Got way over reinstalling and reupdating and reinstalling and reupdating every few months and noticed my 32 bit OS didn't have this problem...
Realy? If i remember correctly microsoft rkeleased 64bit edition just when intel released their first 64bit processor. Microsoft hold release of ready win version more then a year after amd released athlon64
How is it possible that Microsoft wouldn't know/fix this already? Sounds like someone at intel gave a good wad of cash to postpone fixing this 'bug'/feature
We have known for a long time that windows has problems with many cores - more cores, for a long time check 4 core perfprmance under XP, its a known problem with windows for years.
Brilliant video again Wendell
Definitely looks like a bug. Maybe now re-running some gaming benchmarks would be in order. Since UMA could not be enabled on 2990WX, the only "fix" was to use the Legacy Mode. Would be nice to see if this is something CorePrio could work around.
Awesome and Brilliant findings Wendell 👍
You have embarked on a challenging journey and have come out a hero! 💪
Thank you for shining light on such a difficult puzzle 🙏
This is an awesome piece of research and a MUST-WATCH! Respect to all of you
Earned a subscriber because of how analytic and well-paced this was. Keep it up man!
Seema like a conspiracy to nerf 2990wx. You said 100% cpu usage but low performance. The cpu is busy shuffling load on threads instead dedicating fully to the program. Seems done deliberately.
Awesome work! I love spending time on an issue like this and finally getting a solution that makes sense! Congrats!
I take my hat off to you Wendell and confer a healthy dose of virtual applause. Well done.
Awesome Video Wendell! This is quite fascinating and I look forward to more videos on this matter. Thank you for all your hard work.
Now I feel silly for sidelining the 2990wx completely due to the issues I read and watched about earlier and taking them as fact only to see your video now. I'm glad to know it isn't a hardware bottleneck and is an underlying Windows issue. Hopefully they have a fix sometime soon!
ok what did you go with instead of the 2990wx
2950x
@@RISCGames I am so sorry
You have done a great job. You seem to have put a lot of time in this. Your performance score on the 2990WX are the same than mine on Linux 4.19 with PBO enabled on the Threadripper.
Truly Brilliant, well above my pay grade and just another affirmation of my love for this channel being justified.
Much of what I watch on this channel is beyond my meager abilities and comprehension but it's always laid out in a way that I find I can understand and learn from. Reviews are always spot on in my experience and I have never regretted a purchase I've made based on the information provided here. So thank you and stop selling yourself short. You truly are brilliant and a voice that is very much needed/appreciated...
man i LOVE This channel, proper tech lovers
most tech channels are sellouts, especially the big ones & we all know who those are
Great insights. Topology-aware scheduling indeed needs a lot of work but can unlock quite a bit of performance.
Wendell good job, I have been holding back looking into threadripper because of performance drop offs. Come on Zen 2 be all that the rumors are saying and summer of 2019 will be full of cores.
Reminds me of mining Monero with a CPU. You can have Windows hammer the processor, but it results in crap performance. If you manually select physical cores, with specific 'overflow' cores, your hashrates go through the roof. Clearly Windows needs to understand how to address physical cores first, and overflow to the first physical core which started the process AFTER all physical cores are under load.
Thank you Wendel ever so much m8. Your the best. I'm proud to support you.
Awesome video! Great job on describing what is happening. I love these types of videos.
Wendell & Bitsum doing M$'s work!
I've really appreciated your coverage of threadripper, NUMA, and processor topology in general. It has really helped me understand multi-socket topology and scheduling.
About a year ago I built a 1950X system to game on while running a domain controller, pfsense, and other VMs in the background via Hyper-V. Until I saw your videos, I never understood the performance issues I was seeing in games where my GPUs were going significantly underutilized. I knew AMD's IPC isn't up to Intel's, but it never occurred to me that the Windows scheduler might be assigning the threads to the processor not directly connected to the memory channels housing the game data, or that it could be running the threads on a processor not directly connected to the PCIE lanes of the GPU doing the rendering. Your testing shows that the memory latency really isn't that problematic, but I do find it interesting that AMD's initial game mode solution was to disable a die entirely. I didn't realize that the 1950X should really be thought of as 2 8-core processors with dual channel memory, rather than a single 16 core with quad channel. Once I got back into World of Warcraft, I wound up building a new system with an i5-8600K because its old engine strongly favors single thread performance which made the problem even more obvious. I repurposed the 1950X system as an ESXi host where it has done well and even built another to replace my 2 aging and power hungry Dell R610s.
Keep up the great content. You would think AMD and Microsoft would be all over this given it is also affecting EPYC, especially for Hyper-V.
Nice job. I love videos like this. Keep going Wendell. This is the stuff that really keeps me here.
Epic video about Thread-Epic research, love you work guys ^^
I like how Wendell explains it in a way that allows me, with only basic computer knowledge, to understand this.
Finally, properly working boost on 3900X.
Before CorePrio and Bitsum DLM, 4.1-4.2, now, cycles like it should on light loads, +4.5Ghz.
Oh and automated Powerplan switching in ParkControl, man, so good.
This makes perfect sense. On a monolithic die, like the Core series, copying thread context between cores is cheap, so you don't notice much of a performance hit when moving processes from core to another. On a modular design, like TR or Ryzen, copying thread context (stack pointers, registers, etc) is way more expensive when you have to migrate from one CCX to the other, and if that happens a lot, the kernel will thrash. This explains why changing the affinity after process creation fixes the problem, while changing it before process creation doesn't. On top of this is the fact that Windows processes are way more heavyweight than Linux processes anyway, so a context switch would be much more expensive.
One test that might be very interesting to run would be to force benchmarks to run on CCX's that have to cross the IF to get to the memory. In that case, there would be 16 cores, roughly equivalent to a 2950x. If the performance really was similar to a 2950x, you would have further proof that the problem is with the scheduler, not the bandwidth.
You might even try spinning up VM's that use 2 CCX's, but not all the cores; for example, you could have a VMWare instance that uses cores 4 - 11 and compare it to a VM running on cores 0 - 7. I'd bet money that a VM running on CPU 0 - 7 is 50% faster than on 4 - 11 running the same workload, even in isolation.
the problem with this and multithreading in general is that the potential for insane increases in performance is there....but it doesnt always work and to get the cases where it does work to, well, work you have to do some extra work. intels approach is more conservative but it works more often with less work. you just have to decide which path you want to take and if you are willing and have the knowledge to make the amd approach work. the memory bandwidth was overstated and thats a good catch. great video. but youve eliminated an easy problem to fix (just have more mem channels) and found a problem thats more fundamental and harder to fix
SOOOOO Glad this video is up. Bought one a week ago and have been pulling out my hair with a lot of WTF moments until I saw this.
Great work to all involved. Very interesting. Logic at work is enjoyable to watch or read.
good call on saying "U M A" instead of "ooma", i could see myself easily mishearing that as "newma"
I love videos with people that know what they are talking about
This is the stuff that makes this channel great. Great work Wendell.
This is excellent! Great job!
There are smart people in regards to computers etc,etc. Then there is Wendell. Great vid keep up the great work.
Fantastic work! You guys deserve an award, and some prize money or something.
Are you in contact with the Windows Kernel team? Or AMD? I think they might have an interest in your findings.
Known for a while hardware is being developed faster than software this is going to create a problem.
We are seeing it all over the place with patches increasing performance on given hardware ( vega64/56, 10 series gtx cards , amd multicore chips. ) all are seeing year after year increases in performance. Save bloat ware and data harvesting software.
@@texasdeeslinglead2401 Yeah, that new fangled multi-threading gimmick M$ thinks will soon pass. We'll be back to 4 cores on one die.. any day now.
Microsoft seems to be big on spyware and update crashes but small on fixing its own bugs and compatibility. Enjoyed your detailed explanation, even though I don't understand the whole thing. I have also heard that windows have still not fixed their task scheduler since 2009.
Great stuff like always, thanks for shedding some light on this!
It's great seeing you have fun digging into stuff like this :) I feel my Patreon dollars are being spent well!
You should do another video that focuses more so on Process Lasslo and Threadripper, and less Indigo.
I don't use Indigo to stress test the 2990wx. I use multiple Hyper-Vs that I have loaded in Hyper-V Manager in Windows 10 Pro for Workstation. Each Hyper-V is assigned 4 Virtual Processors. On each VM I run a CPU miner for mining Monero on each VM. I use 64GB of RAM.
When I initially set up my Threadripper workstation, the Windows Task Manager operated fine and it showed that it was gradually utilizing CPU when I increased the amount of VMs.
3 months in and my system completely crashes, causing me to use the Reset feature in Windows Recovery. After that the Task Manager would always show 100% Utilization even though I was using less VMs than I did prior to the crash.
Recently I installed Process Lasslo and now I am seeing proper thread management and CPU utilization. I am now running all my Hyper-Vs at once 14 totaled. In addition I have 3 miners on the host, 2 GPU mining etherium, 1 dual-FPGA mining zen protocol.
-----> A minor animation error to look at: In the logo (the bottom right corner if it) from 0:20 - 0:23 there is the last piece of unveiling logo - not unveiled.
This is not a critic of your video, but I noticed it at the beginning of playing. I like your videos, BTW. :)
Phenomenal work guys. Wendell, Jeremy, et al.
i just installed coreprio and even on my plain jane ryzen 7 2700 with a gigabyte mini gtx1060 and I got a huge performance boost. my fps in cinebench opengl went from 97.97 to 121.63. my cpu score went from 1473 to 1514. I am using exactly the same hardware as I did to achieve those previous scores.
I just hope that AMD is prodding Microsoft to get this fixed. We should all petition Microsoft to fix this ASAP.
I would suggest to start upvoting that issue on Feedback HUB in Windows 10.
The whiteboard breakdown was great, more of that good sir.
Fuck yes, Wendell talking about CPU architecture. I love videos like this!
I've never seen a like/dislike ratio like this. After watching this video I realized why. Keep up the good work! :)
Very interesting article. What specific OSs did you test with? Did you try workstation and server builds too?
Yup right on the nail!
Windows kernel at a lower level.....was not built for TR architecture. Then again when windows was coded, the coders never ever thought any CPU manufacturer would create something like the TR.
I guess the Linux camp had a different thought when it came to supporting single CPUs with multiple cores.