Linux performance crushes Windows on new 32-core AMD flagship

[Humbug]

The release of the new generation of Threadripper CPUs has been met with widespread acclaim for the 16 core 2950x which performed well on any workload, but more mixed results for the 32 core 2990wx which seemed to

a) not scale quite as well as expected on some multi-threaded applications (even when it is the fastest CPU around)

b) perform lower than expected on memory bandwidth sensitive applications

 

Phoronix has conducted a plethora of multi-threaded tests using multi-platform applications that are available for both Linux and Windows. The tests compared Windows 10 performance to four Linux distros- Ubuntu, Clear Linux, Antergos and OpenSuse.  

 

The current speculation is that Windows is not as well tuned as Linux to NUMA (non-uniform memory access). But we are yet to hear any official word from AMD or Microsoft, so nothing concrete.

 

Still the 2990x is a memory bandwidth starved CPU as many have observed. So it may be a case of things getting compounded on windows by being both memory starved and not optimized for NUMA.

 

Click on below image for full res version

 

Quote

The Windows 10 vs. Linux tests were done out of opportunity with having that Windows installation around without giving it too much thought, but on this 2990WX launch-day it's been surprising to see some of the performance results from some of the Windows publications. Had I known how poorly Windows 10 works on current high core count NUMA environments under some workloads, I would have certainly ran more benchmarks. But that will come in another article then as well as possibly looking at the Windows Server 2016 vs. Linux performance on the 2990WX to see if Windows behaves better there for this NUMA box. So treat this as the introductory article and more Windows vs. Linux benchmarks will be on the way as time allows.

 

Throughout this benchmarking the hardware was maintained the same of using the Threadripper 2990WX at stock speeds with the ASUS ROG ZENITH EXTREME motherboard, Cooler Master Wraith Ripper heatsink, 4 x 8GB G-SKILL DDR4-3200MHz memory, 500GB Samsung 970 EVO NVMe SSD, and Radeon RX Vega 56 graphics.

-With the Stockfish multi-threaded chess engine, Intel's Clear Linux was performing in line with Windows 10 Pro at the bottom of the fact while openSUSE Tumbleweed maintained a slight advantage over the other tested Linux distributions.
-With the chess benchmark, openSUSE was 20% faster than Windows 10 on this Threadripper 2990WX system. The Linux distributions did come out ahead as well in the Primesieve prime number benchmark.
-In basic single-threaded test cases like FLAC audio encoding and BLAKE2 where there is good cross-platform support, the performance comes out to being about the same between Windows and Linux for this 32-core high-end desktop processor.
-Clear Linux and Windows 10 were performing the best when it came to a basic FFmpeg video encode benchmark.
-With the OpenMP-based M-Queens (an N-Queens solver), the performance between the five tested operating systems were very close.
-Within the Fhourstones CPU benchmark, Windows 10 was the slowest but in this benchmark the spread isn't nearly as wide as some of the other tests... OpenSUSE Tumbleweed was the fastest in this test and came out to being 35% faster than Windows 10. Interestingly Clear Linux in this case was the slowest Linux distribution tested.
-The Crafty multi-threaded chess engine was another one of the cross-platform benchmarks where the spread wasn't all that great: OpenSUSE Tumbleweed again was the fastest but only 10% faster than Windows.
-The x264 video encoding performance was notably faster on Linux than Windows 10 at least on the platforms where there wasn't build configuration issues.
-We've seen it with other hardware too, but the GraphicsMagick imaging program that relies upon OpenMP for multi-threading does perform significantly better with Linux over Windows. But then again Windows users have Adobe products...
-With the 7-Zip compression benchmark, the Linux distributions all perform significantly better than the Windows 10 Pro result. (And our Windows 7-Zip result does appear to be in line with what some of the Windows publications reported as their performance.) The Linux distributions in this extreme case were about twice as fast.
-Linux tended to perform faster than Windows with the Minion constraint solver, but the Microsoft OS did come out ahead in the Solitaire sub-test.
-Blender is one of the fun tests that can really pound the 2990WX across all 64 threads. With this pts/blender test profile the official Linux/Windows binaries from Blender.org are what are automatically used for benchmarking. Here we see all four tested Linux distributions offering significantly better rendering times than the Windows performance. (These Windows numbers also jive with what was seen when checking a few Windows publications where the same Blender scenes were used.)

 Long story short, the Linux performance in a majority of these CPU-focused benchmarks were running much faster on the AMD Threadripper 2990WX than Windows 10 Pro when tested with the same hardware in the same configuration. Then again, we usually see better performance with Linux over Windows on most hardware but not always to some of the extremes encountered. It will certainly be interesting to run more Windows vs. Linux tests on the 2990WX Threadripper platform moving forward.

 

https://www.phoronix.com/scan.php?page=article&item=2990wx-linux-windows&num=1

[Bananasplit_00]

I wouldn't be surprised if Windows server sees better performance numbers, I wouldn't expect any enormous improvement but I'd think it might catch up in some benchmarks. It seems awfully far behind in blender when run on Windows compared to Linux though

[cj09beira]

1 hour ago, Bananasplit_00 said:

I wouldn't be surprised if Windows server sees better performance numbers, I wouldn't expect any enormous improvement but I'd think it might catch up in some benchmarks. It seems awfully far behind in blender when run on Windows compared to Linux though

i bet it wont make a difference 

[ScratchCat]

Part of this as mentioned in the other thread will have to do with the strange NUMA architecture of Threadripper. Almost all normal NUMA systems have a pool of memory per core cluster, be the cluster a Zen die or a full CPU. 

 

TR2 uses a different approach where two dies are connected to the memory directly while the other two dies have to hop to memory over the IF incurring additional latency which, as mentioned in the review thread, will cost TR2 dearly in tests such as decompression. 

 

AMD did plan to prioritise cores with direct memory access over the others however with some workloads this will not be possible. In this case the probably TR2 aware 4.19 kernel may be able to compensate for this (somehow).

[Humbug]

9 hours ago, ScratchCat said:

AMD did plan to prioritise cores with direct memory access over the others however with some workloads this will not be possible. In this case the probably TR2 aware 4.19 kernel may be able to compensate for this (somehow).

If it can it's a shame that they couldn't sort this out with Microsoft as well before launch... Obviously more difficult than dealing with the transparent Linux Kernel but still.

[leadeater]

Well I'm surprised there's so little comments in this topic, I was expecting it to be at like page 9 but now .

[Humbug]

4 minutes ago, leadeater said:

Well I'm surprised there's so little comments in this topic, I was expecting it to be at like page 9 but now .

Maybe I should have made a more clickbaity title. e.g.

 

"Linux performance crushes Windows on new 32-core AMD flagship"

[The Benjamins]

14 minutes ago, leadeater said:

Well I'm surprised there's so little comments in this topic, I was expecting it to be at like page 9 but now .

Their was 20+ post in the repost thread that was locked.

 

[leadeater]

1 minute ago, The Benjamins said:

Their was 20+ post in the repost thread that was locked.

Only 2 of the 9, 7 more to go . But yea I commented in that one, not much to say other than way more testing required.

[yolosnail]

11 minutes ago, Humbug said:

Maybe I should have made a more clickbaity title. e.g.

 

"Linux performance crushes Windows on new 32-core AMD flagship"

It's not clickbaity if it's factually accurate, which it is

[Humbug]

32 minutes ago, yolosnail said:

It's not clickbaity if it's factually accurate, which it is

ok, point taken

[PlayStation 2]

Those Blender results, damn.

I'm also super curious on the details of the x264 encoding.

[yolosnail]

Just now, Humbug said:

ok, point taken

Now watch the engagement roll in

[Taf the Ghost]

6 minutes ago, Dan Castellaneta said:

Those Blender results, damn.

I'm also super curious on the details of the x264 encoding.

It's going to depend on what's being used to encode with, and what flags are set. Encoding is its own art form.

[leadeater]

1 hour ago, yolosnail said:

It's not clickbaity if it's factually accurate, which it is

Umm, actually...

 

 

Spoiler

Made you look

 

[Doobeedoo]

I guess some further optimization fix is needed though.

But it also seems that for memory focused tasks where CPU needs info most of the time from memory it needs to hop back and forth. Thus extra latency and now that there are more than two dies active too. So also it's cause of nature of MCM design maybe.

[yolosnail]

Just now, leadeater said:

Umm, actually...

 

 

  Reveal hidden contents

Made you look

 

Now that's clickbait!

[mynameisjuan]

1 hour ago, leadeater said:

Well I'm surprised there's so little comments in this topic, I was expecting it to be at like page 9 but now .

No its would be 9 pages if it was "MacOS performance crushes Windows on new 32-core AMD flagship"

[AngryBeaver]

I think the reason it isn't on page 9 is that most people can see this for what it is worth. Linux has better support for a completely new chip, one with a funky design. So that means the performance windows is currently seeing is due to a compatibility issue and will be straightened out in quick order via a patch or hotfix. So while linux has a edge currently, I fully expect that to disappear in a few weeks or so when Microsoft releases their fix.

 

That is like comparing a GPU with beta drivers to full release drives 2 or 3 revisions later. It isn't exactly a fair comparison. I mean one of the strengths of linux has always been the compatibility with hardware.

[Taf the Ghost]

7 minutes ago, AngryBeaver said:

I think the reason it isn't on page 9 is that most people can see this for what it is worth. Linux has better support for a completely new chip, one with a funky design. So that means the performance windows is currently seeing is due to a compatibility issue and will be straightened out in quick order via a patch or hotfix. So while linux has a edge currently, I fully expect that to disappear in a few weeks or so when Microsoft releases their fix.

 

That is like comparing a GPU with beta drivers to full release drives 2 or 3 revisions later. It isn't exactly a fair comparison. I mean one of the strengths of linux has always been the compatibility with hardware.

I also think it's summer in the USA and people are just less inclined to argue about those things. Come Fall and colleges being back in session..

[Humbug]

57 minutes ago, AngryBeaver said:

I think the reason it isn't on page 9 is that most people can see this for what it is worth. Linux has better support for a completely new chip, one with a funky design.So that means the performance windows is currently seeing is due to a compatibility issue and will be straightened out in quick order via a patch or hotfix. 

In that case AMD should have mentioned it, at least in the reviewer's guide.

[AngryBeaver]

3 minutes ago, Humbug said:

In that case AMD should have mentioned it, at least in the reviewer's guide.

For all we know they already have a beta patch or similar they are supplying to reviewers. I mean do you think they demo'd this thing and/or benchmarked it without having a way to correct that issue. Chances are the patch already exists in a beta stage... it just isn't ready for release yet.

[rcmaehl]

Clickbait is just a product journalism. Similar to yellow journalism in the past, consumers only have limited interest and time to spend reading information, all information sources are competing for a persons time and attention. I don't mind factual clickbait titles, but vague uninformative clickbait titles such as "you won't believe what this 12 year old did" and "Is Elon Musk converting the Model 4 into a Hybrid?" I absolutely despise.

[cj09beira]

11 minutes ago, rcmaehl said:

Clickbait is just a product journalism. Similar to yellow journalism in the past, consumers only have limited interest and time to spend reading information, all information sources are competing for a persons time and attention. I don't mind factual clickbait titles, but vague uninformative clickbait titles such as "you won't believe what this 12 year old did" and "Is Elon Musk converting the Model 4 into a Hybrid?" I absolutely despise.

click bait is the product of lack of ethics and poor journalism nothing more nothing less

[TVwazhere]

5 hours ago, leadeater said:

Umm, actually...

  Hide contents

Made you look