New GeForce Driver Fixes Threadripper 2990WX Performance

[vitor_cut]

It seems like the green team, was able to extract some impressive improvemnets on the 32 core Threadripper.

Mr. Ryan Shrout retested the 2990 WX and saw some interesting numbers.

In my opinion more instances as this one will happen, but the limited bandwidth inter CCX will be always felt in aplications as games.

It would be sweet if the rumours about X499 with 8 memory chanels turns to be true, it would be very nice for me for solidworks work load and CAM programing ,but i dont have much hope.

 

Spoiler

Quote

Earlier today, NVIDIA released version 399.24 of their GeForce drivers for Windows, citing Game Ready support for some newly released games including Shadow of the Tomb Raider, The Call of Duty: Black Ops 4 Blackout Beta, and Assetto Corsa Competizione early access. 

 

 

While this in and of itself is a normal event, we shortly started to get some tips from readers about an interesting bug fix found in NVIDIA's release notes for this specific driver revision.

 

 

Specifically addressing performance differences between 16-core/32-thread processors and 32-core/64-thread processors, this patched issue immediately rang true of our experiences benchmarking the AMD Ryzen Threadripper 2990WX back in August, where we saw some games resulting in frames rates around 50% slower than the 16-core Threadripper 2950X. 

This particular patch note lead us to update out Ryzen Threadripper 2990WX test platform to this latest NVIDIA driver release and see if there were any noticeable changes in performance.

The full testbed configuration is listed below:

Test System Setup
CPU	AMD Ryzen Threadripper 2990WX
Motherboard	ASUS ROG Zenith Extreme - BIOS 1304
Memory	16GB Corsair Vengeance DDR4-3200 Operating at DDR4-2933
Storage	Corsair Neutron XTi 480 SSD
Sound Card	On-board
Graphics Card	NVIDIA GeForce GTX 1080 Ti 11GB
Graphics Drivers	NVIDIA 398.26 and 399.24
Power Supply	Corsair RM1000x
Operating System	Windows 10 Pro x64 RS4 (17134.165)

Included at the end of this article are the full results from our entire suite of game benchmarks from our CPU testbed, but first, let's take a look at some of the games that provided particularly bad issues with the 2990WX previously.

The interesting data points for this testing are the 2990WX scores across both the driver revision we tested across every CPU, 398.26, as well as the results from the 1/4 core compatibility mode, and the Ryzen Threadripper 2950X. From the wording of the patch notes, we would expect gaming performance between the 16-core 2950X and the 32-core 2990WX to be very similar.

Grand Theft Auto V

 

 

GTA V was previously one of the worst offenders in our original 2990WX testing, with the frame rate almost halving compared to the 2950X.

However, with the newest GeForce driver update, we see this gap shrinking to around a 20% difference.

Continue reading our revised look at Threadripper 2990WX gaming performance!!

Assassin's Creed: Origins

 

 

While Assassin's Creed: Origins remained playable on the 2990WX with a frame rate of 58 FPS, there was still a significant performance gap of 43% between it and the 2950X.

With the new driver, this performance gap shrinks to 13% and brings the 2990WX in its default 32-core mode to the same performance levels that we saw in the 1/4 core legacy compatibility mode initially.

Total War: Warhammer II (DX11)

Total War: Warhammer II was far and away the biggest presenter of this performance issues, going from frame rates in the 70s-90s with every other processor we tested, down to average frame rates in the teens.

 

 

While we see a strong performance uptick of 25-35% with the new driver, the resultant frame rates still don't provide a playable experience.

F1 2017

 

 

Updating to the latest NVIDIA driver provides a massive 78% improvement for the Threadripper 2990WX, and closes the gap to just 7% between the 16-core and 32-core processors.

Middle-Earth: Shadow of War

 

 

Previously, Shadow of War showed high average frame rate scores but was still provided an almost 20% deficit when comparing the 2990WX to the 2950X.

With 399.24, this performance gap disappears entirely, providing a 30% performance increase when pairing a GTX 1080 Ti and the Threadripper 2990WX.

Far Cry 5

Far Cry 5 is an interesting title because, during our initial review of the 2990WX, we found that the game was incompatible with the 32-core processor and crashed on demand. However, be it due to a game patch or driver changes, it ran fine on our 2990WX-equipped system with both of these updates.

 

 

Despite the newly added compatibility, the 2990WX still sees half the frame rate in 32-core mode versus 1/4 core compatibility mode.

 

 

Overall, we are quite impressed with the performance increase this new GeForce driver provides over previous drivers when it comes to the 32-core Ryzen Threadripper 2990WX. Almost every title we tested saw some sort of frame rate increase, up to 78% for the most improved title, F1 2017.

However, there are still some caveats to recommending the Ryzen Threadripper 2990WX to content creators who are also interested in gaming. We saw some heavy stuttering while testing in a few titles, namely Assassin's Creed: Origins and F1 2017 that was not present while testing the same platform in 1/4 core mode. We've passed this data onto NVIDIA and hope they will continue to refine how their driver handles these extremely high-core count processors.

 

 

In general, the architectural decisions made, specifically in the memory department, that make the 2990WX a fantastic workstation processor do inherently still hold it back in gaming, when compared to other processors and even the Threadripper 2950X. For users who are going to run games anyway on this processor, it's nice to see that this new driver has eliminated the need to reboot into Legacy Compatibility mode for the vast majority of games.

To reiterate what AMD has been saying about the 2990WX from the beginning - this part isn't INTENDED for gamers. That's fine, we totally get it. But we like to provide the data that the audience is looking for and gaming results will at least be part of that.

Still, we are heartened to see such an improvement in game performance for the 2990WX, and these driver-level changes from NVIDIA should apply to other processors in the future like AMD's upcoming 24-core 2nd generation Threadripper processor and Intel's announced 28-core desktop offering.

Ryan's Note: This is an interesting shift in performance that lends itself to a deeper discussion at some point in the future. First, if this was indeed just a bug in the NVIDIA driver itself, the timing of it could not have been worse for AMD, as the window of time it existed is very closely matched with the release of the 2nd Generation Threadripper reviews. Second, this does make us alter our view of the Threadripper 2990WX. As Ken noted, the processor still has complications for gaming caused by its architecture but the picture is improved dramatically with these updated results. 

It also reveals a shortcoming of our gaming testing, and the gaming testing of most of the online community - we only tested with a single graphics card. We use a GeForce GTX 1080 Ti for our CPU reviews because NVIDIA has been known to have better performance and better driver stability, but this situation has us a bit concerned. Best case is we double our workload and test a Radeon graphics card too. But then that can waterfall to different architectures (why not Turing AND Pascal, or Vega AND Polaris?), different driver versions, etc. It's a tough battle for us to face.

For AMD itself, this is a learning opportunity. The company said nothing to us to make us second guess our results, or push us to do Vega-based testing that would have shown the 2990WX in a more positive light. If the engineers were doing in-line testing for some longer period of time during product development, they would have seen this sudden drop in 2990WX gaming performance with a specific NVIDIA driver drop, and know to address it with NVIDIA or the media. Instead, it seems that everyone involved was in the dark. (Though AMD was leaning on the WX-line as a workstation only part, it surely was doing some kind of gaming testing.) Maybe NVIDIA knew about the bug - but rather than inform the community, it decided to let the reviews of the AMD Threadripper processors go out unfettered.

The whole situation has really been a mess.

 

Quote

To reiterate what AMD has been saying about the 2990WX from the beginning - this part isn't INTENDED for gamers. That's fine, we totally get it. But we like to provide the data that the audience is looking for and gaming results will at least be part of that.

Still, we are heartened to see such an improvement in game performance for the 2990WX, and these driver-level changes from NVIDIA should apply to other processors in the future like AMD's upcoming 24-core 2nd generation Threadripper processor and Intel's announced 28-core desktop offering

 

source: https://www.pcper.com/reviews/Graphics-Cards/New-GeForce-Driver-Fixes-Performance-Issue-Threadripper-2990WX-Tested

[PlayStation 2]

What the actual hell is going on with these frame rates?

[vitor_cut]

Its the price of that unusual CPU arrengment that AMD implemented in Threadripper WX series to protect their Epyc line.

[Trixanity]

FineWine™

[AlTech]

42 minutes ago, Dan Castellaneta said:

What the actual hell is going on with these frame rates?

Nvidia drivers have been known to screw up Ryzen and Threadripper performance.

[MMKing]

Thats cool and all, but i don't find much value in benchmarking a 1750USD CPU at 1080p. Who is this person that spends 1750USD on a CPU, another 200-400USD on memory, an expensive 12-16 phase power motherboard, a beefy high quality power supply, presumably an SSD and a sizable case to put it all in, then only puts up 350-450USD total for a GPU + Monitor?

[Taf the Ghost]

1 hour ago, Dan Castellaneta said:

What the actual hell is going on with these frame rates?

There's two things:

 

1) Too many cores makes certain games spaz out. This happened with Ryzen, Threadripper 1 and now Threadripper 2 releases. Expect it again with Threadripper 3.

2) Nvidia's driver does not like the non-monolithic designs. This has been true since Ryzen launched. (It has more to do with the non-exclusive L3 cache as much as anything, as X299 had some real trouble when it first launched that Nvidia smoothed out in a few months.)

 

Some games are still buggy with that many threads being active. (A few games saw normal results with SMT off, so it was >36 threads that did them in.)

[Taf the Ghost]

9 minutes ago, MMKing said:

Thats cool and all, but i don't find much value in benchmarking a 1750USD CPU at 1080p. Who is this person that spends 1750USD on a CPU, another 200-400USD on memory, an expensive 12-16 phase power motherboard, a beefy high quality power supply, presumably an SSD and a sizable case to put it all in, then only puts up 350-450USD total for a GPU + Monitor?

Pretty much every game is GPU limited at 4K on any of the recent processors. 1080p testing at least lets you see how the clocks & memory setup are playing along together. And 1080p high refresh gaming is a thing, so 1080p testing is still valid. 720p testing on the other hand...

[Curufinwe_wins]

56 minutes ago, AluminiumTech said:

Nvidia drivers have been known to screw up Ryzen and Threadripper performance.

Actually... it's just what happens when non-numa aware software works on a CPU that implicitly requires NUMA programming. 

 

And it's a bad CPU design for a hyper majority of current computation environments, even ignoring games. Just straight up bad.

[AlTech]

1 minute ago, Curufinwe_wins said:

Actually... it's just what happens when non-numa aware software works on a CPU that implicitly requires NUMA programming. 

 

And it's a bad CPU design. Just straight up bad.

Ryzen doesn't use NUMA. And yet Nvidia has had this issue in the past with Ryzen.

 

Threadripper does and so it's a bit of a different animal.

[Curufinwe_wins]

8 minutes ago, AluminiumTech said:

Ryzen doesn't use NUMA. And yet Nvidia has had this issue in the past with Ryzen.

 

Threadripper does and so it's a bit of a different animal.

Nvidia AND AMD consistently have issues with difference cpu designs from the same manufacturer. Nvidia had to roll out driver updates for Skylake-X for similar cache design differences that caused Ryzen issues. 

 

Your comparison between "growing pains" and the threadripper problems (particularly the 4 module, 2 module fully functional sku(s)) is invalid. 

[Taf the Ghost]

10 minutes ago, Curufinwe_wins said:

Actually... it's just what happens when non-numa aware software works on a CPU that implicitly requires NUMA programming. 

 

And it's a bad CPU design for a hyper majority of current computation environments, even ignoring games. Just straight up bad.

It's not the NUMA. It's the slightly imbalanced cache & memory latencies. Nvidia has improved the situation, but their driver was pretty much designed to handle the i5s and not scale well beyond that. What AMD ended up doing was showing the weak points in Nvidia's driver design. 

 

Considering Intel has been talking chiplet for quite a while (and that's where the industry is headed), it's not like it wasn't eventually going to be an issue for them.

[Taf the Ghost]

The NUMA stuff is on Windows' Scheduler. That's their area. Recognizing thread counts is on Nvidia.

[AlTech]

7 minutes ago, Curufinwe_wins said:

Nvidia AND AMD consistently have issues with difference cpu designs from the same manufacturer. Nvidia had to roll out driver updates for Skylake-X for similar cache design differences that caused Ryzen issues. 

 

Your comparison between "growing pains" and the threadripper problems (particularly the 4 module, 2 module fully functional sku(s)) is invalid. 

Modern AMD CPUs don't have modules. They have CCXs and CCXs don't function like modules.

 

Each CCX has 4 cores and 8MB of L3 cache.

 

CCX to CCX communication has a bit of latency but it's not bad. Threadripper 1st gen and Threadripper 2nd gen Up To the 2950X is basically 2 Ryzen dies on the same chip. And so in that case there is NUMA but it's not a bad system.

 

Threadripper 2970WX and 2990WX have the same issues as previous Threadripper but on a larger scale since it effectively uses 4 Ryzen dies on the same chip.

 

Most developers aren't familiar with NUMA yet and so mediocre or subpar performance is expected in some workloads where CPUs have NUMA such as the Threadripper 2990WX.

 

3 minutes ago, Taf the Ghost said:

The NUMA stuff is on Windows' Scheduler. That's their area. Recognizing thread counts is on Nvidia.

Yeah.

[leadeater]

22 minutes ago, AluminiumTech said:

Threadripper 2970WX and 2990WX have the same issues as previous Threadripper but on a larger scale since it effectively uses 4 Ryzen dies on the same chip.

 

Most developers aren't familiar with NUMA yet and so mediocre or subpar performance is expected in some workloads where CPUs have NUMA such as the Threadripper 2990WX.

There is still a problem with the TR2 4 die SKUs, NUMA is actually about memory boundaries not chiplets or dies and a NUMA node is supposed to have memory but 2 of the dies don't have memory. You've got 2 null NUMA zones basically which is not an assumption a traditional NUMA application is going to make, all NUMA zones are supposed to be equal and they are not.

 

Linux patched the Kernel to add awareness for this imbalance but I doubt Windows has yet, or ever will. I don't see this 4 die 2 memory zone architecture living past this current socket generation.

[Taf the Ghost]

16 minutes ago, Tedny said:

He is more about who will even thinking to us threadripper for gaming 

Well, it's a unique platform because of the Nodes, so it's interesting testing anyway. And people do game on them, as their high-end enthusiast platforms as well

[S w a t s o n]

3 hours ago, Taf the Ghost said:

Pretty much every game is GPU limited at 4K on any of the recent processors. 1080p testing at least lets you see how the clocks & memory setup are playing along together. And 1080p high refresh gaming is a thing, so 1080p testing is still valid. 720p testing on the other hand...

lets be honest, in terms of overall actual population 1080p 60hz is still the norm by a wide margin so 1080p testing is still the MOST valid technically