AMD Ryzen Threadripper 2990WX performance can see up-to 2X Boost with CorePrio tool

Yes, separately Wendell confirmed Windows Server uses the same flawed kernel.

 

I am no chip designer, and no one is more solidly on the red team.......but looking at the circuit diagram for the threadrippers, giving only two of the four ccx direct memory pathways seems like a mistake. Maybe they couldn't, or maybe they had their reasons, but it seems like quite a bottle neck. If someone wants to send me a 2990wx and motherboard (maybe a couple titans, 128 gigs of ddr4 3200, about 4 2tb 970 pros, and an nvme raid host card, ya know....the basics) i would be happy to test CorePrio and report back with my results.....totally won't gank all the fancy gear

 

I'm not sure what your logic is. If you're suggesting that a multi-billion dollar company shouldn't be collecting their customer's data for bug tracking and internal statistics, then why does free software not usually do this sort of data collection?

You might not have an obligation to them (I don't either), but a lot of people who need to run Windows-compatible software do. But, I do for the most part agree with the gist of your statement, though to extrapolate further: if you buy hardware for a certain version of a piece of software that you also bought, they should work correctly without any fuss, and the 2990WX was not one of such cases.

Yup, pretty much.

Actually that's more because of how lightweight and customizable Linux is, but I digress.

 

When Threadripper launched, I saw on friends machines that load balancing was not respecting CCX package barriers, I concluded that tagging was not working properly at that point.

Threadripper 2 comes along and I said "this is going to hurt", microsoft still hadn't corrected the tagging, so load balancing behaviors will juggle threads that should remain on one package between 2 or more.

The result?, Cache invalidation and branch misprediction penalties, when a task was moved from one package to another everything had to drop back to the uops cache and the high utilisation as has been seen was processing being done all over again.

non-localised memory on the detached CCX's are not the issue. See above.

 

My logic was that I had to pay for software, QA is nowadays normal job and hence I produce product in form of data and should be paid for it , same way as I did paid for Windows.

My comment was jab at multi billion company that's unable to resolve issues on their own. It's not like these new CPU come to the market overnight, they had time to develop and adopt (including my i9 that I have issue with).

As stated before my, or Epyc / TR WX, setups are pretty much standard and they work on Linux more or less fine, so where is M$ QA? Looking back at two last updates, driver issues and user file deletion, I wonder what's next?

This was regarding their non existent QA. Nowadays, to me, it seams like users are QA.
As for CPU itself it should work correctly and not shut down, as it seams, half CPU cores (excluding XFR and all new features)

You can't say that actually as we don't know how Windows kernel looks like on src side .
Personally I think it's all about money not about "innovating" or making better products. Same shi3t with Minix and Intel now it's M$ and kernel well at least M$ says laud that they will use Lin kernel not like Intel.

 

Thank you ! Its been a few years ,back in my phenom day , he had A tool called LASSO, just reinstalled the latest one on my system, AMD 8370E , testing it out now.

And downloaded power profile, Ultimate performance from his page , to compare.

 

Well they kind of have to. AMD going leaps and bounds with core count increases across the board is something they kind of have to do imo to get a leg up on Intel, price alone won't cut it.

It's great! It'll force Intel to do the same or at the very least cut cost because they'll bleed marketshare, AMD knows Intel has been stagnant with core count for desktop so now they're ripping up the rules so to speak and going all out. I'm guessing Zen2+ (the refresh after Zen2 about to come out) it'll all stay the same with probably higher core clocks, Zen3 will push the cores higher again is my guess.

 

Yeah I replied back to him on it as well. It appears this may be a big find I hope it gathers a ton of attention to force Microsoft to update its kernel as this could be huge for data centers especially as EPYC on Zen 2 lands later this year.

Michael Larabe had already shared Windows server 2019 results and it does appear to have the same issue. Once the threads are over 32 you can clearly see performance degrades in Windows compared to Linux. He didn't try to figure out why just published test results.
https://www.phoronix.com/scan.php?page=article&item=2990wx-linwin-scale&num=4

 

Since 32C/64T TR can do with quad channel memory configuration quite well, I think that 16C/32T will do quite well with dual channel memory configuration.

But above that I am very unsure. It will likely need much faster memory (4GHz+) or latency improvements. (That's if AMD want's to keep AM4 socket.)
And since even Vermeer planned for 2020 is listed for now as AM4, I do not expect new socket which would allow for DDR5 or quad channel on consumer platform.

 

Core Prio should have been done by AMD themselves, not the program but on the actual chip or firmware/drivers.

 

It fixes errors in Windows kernel. All meaning full fixes between kernel and firmware are done other way around.
When firmware (BIOS, ...) has bug, there is fix in kernel. That makes sense, not other way around.

But reading comments in AMD's driver handling those issues like:
# Is microsoft gimping our threading again?

 

It was known windows scheduler has serious issues few days after 1st gen ryzen launch. then AMD mysteriously stated they are sure everything with windows is fine. it was sketchy af imo and now it doubled

 

Looking at AMD's timeline suggests Zen3 will not increase core counts as it is 7nm+. But Zen2 and Zen3 should be PCIE4 compatible. So B550, X570, X599 will be DDR4 with PCIE4 slotted motherboards.

So Zen4 could be new cores on 5nm, DDR5, PCIE4 in new socket. Zen4 is supposed to be 5nm, which is not as large a drop in size compared to Zen+ to Zen2. This is the time for memory channel changes if required. Perhaps the EPYC IO chip will have shrunk by then and will allow for more chiplets, and on desktop triple channel memory will be an option. Lets wait and see what Jim@AdoredTV comes out with.

 

Maybe after they fix this Microsoft can support nested virtualization on AMD processors like everyone else does..

Seems like they are intentionally being obtuse about issues on AMD.

 

Absolutely a great youtube to watch , thanks for sharing.

 

Doubled my 2970WX performance with the Indigo Benchmark as indicated, and other gaming Benchmarks as well; (3DMark, Unigine Superposition, etc...).

 

IMO this will get "fixed" in windows when Intel releases their 28 core CPU.

 

IMO this is AMD racing to bring something sensational to market before it actually was ready. This isn't MS's fault, and this "fix" is lacking...

The Xeon 8180 is released, and it actually works. Intel worked with Microsoft BEFORE the release to ensure that it actually functions as intended...

Phoronix showed yesterday that Coreprio reduced performance in various loads. It can hurt as much as it helps depending on the load. Server 2019 performed very differently than Windows 10 in some of Michael's tests. Indigo does even worse on Server 2019 than on Windows 10, while in other benchmarks Server 2019 beats Windows 10, with or without CorePrio.