I recently purchased an Arc B580, and although I really like how it performs in almost all games I tried, I was quite disappointed by its performance on CS2. The card has a significant performance drop when using DX11 compared to other cards and when compared to using Vulkan as the game's API, which you can opt into by using the launch command "-vulkan". In my tests, using DX11 as the API for the game made the game run 20% slower than Vulkan, and almost 30% slower when Low Latency mode is enabled.
This is not the case with any other card from Nvidia and AMD; using Vulkan on those cards will actually make your performance slightly worse. My fear is that CS2 running on DX11 is such a nightmare for the Arc B580 that using Vulkan will make the card appear to run the game better, but in reality, it's just because the card delivers bad performance on DX11, which maybe is related to a possible CPU overhead issue in CS2 since I saw many people reporting that using Vulkan in other games that presented the overhead problem helped the performance.
My thoughts seem to make sense when compared to my friend's PC, which is identical to mine except for the GPU, as we bought the parts for our PCs together. My old GPU results also seem to indicate the same.
Compared to his 8GB RTX 3060, the B580 was around 22% slower in the scenario I tested, and it only could match the RTX 3060's performance using Vulkan. The problem is that CS2's implementation of Vulkan is not ideal; it's very unstable, with constant crashes and users mentioning latency issues with the game. I agree with those claims; even though Vulkan should be smoother, the game just feels laggy when using Vulkan compared to DX11, which feels better, especially when using the Low Latency mode. So using Vulkan in just not ideal for competitive players.
Apart from comparing the results with my friends 3060, I can also compare it with my old card, the GTX 1650. ( I did use DDU before installing the Arc B580 just to make that clear).
My 1650 gave me 322 average fps on the same settings my friend and I tested (all low, no FSR, 1280x960). The B580 almost looses compared to the 1650 when using Low Latency, which in the test I used to measure the performance with the 1650, Nvidia reflex was enabled. So these are awful results considering the GTX 1650 is a 6 year old low-end GPU, it should be notable better than the RTX 3060. Most of the tests I saw the support team doing were made with the game on ultra/high settings and also at 1080p, which are useless to me or to most competitive players as no one really plays with those settings.
I'd like to ask for this seemingly unexplained performance drop the Arc B580 has with DX11 on CS2 to be investigated further, if possible. I'll attach my full specs, data, explain how I gathered the data presented, and how to reproduce the results.
PC Specs:
- Processor: AMD Ryzen 5 5600X 6-Core Processor.
- Maxsun Intel Arc B580 Milestone 12G (driver version 32.0.101.8136)
Motherboard: TUF GAMING B550M-plus.
Memory: Corsair Vengeance RGB RS 16GB DDR4 3200MHz (2x8).
Storage: Game and OS are installed on a NVMe M.2 drive.
OS: Windows 10.
Note: Rebar is turned on and PCIe is set to 4.0 in the BIOS settings.
Rebar is turned on and PCIe is set to 4.0 in the bios settings.
To make the benchmark averages I used the Steam Workshop map CS2 FPS BENCHMARK DUST2, using CapFrameX to record the average of 3 runs then making an average out of the results. On the end of each run, the map displays the average fps results on the developer console Window. I reached the conclusion that those results are accurate, because they are always really close to the averages CapFrameX registers. Knowing this, I asked my Friend to do 3 runs using specific video settings, then I did an average of the results, and compared said results with mine. His RTX 3060 made an average of 479.5 fps compared to the B580's 375.9 on DX11, making the Arc Card 21.61% slower compared to the 3060.
How to reproduce the results:
Settings used: Fullscreen, 1280x960, 4:3 aspect ratio (it may sound crazy, but most CS players play the game using this video resolution).
Low settings preset, set both FSR and boost player contrast to disabled.
Open the Dust 2 benchmark map using the workshop maps tab, and start capturing the performance as soon as the countdown is over and the game itself shows up. End the capture as soon as the game starts to fade out to display the results.
That should be all. Thank you for your time and attention. I hope I've been informative so far and I look forward to any potential updates in this matter. If any extra info is needed, please, let me know.
Link Copied
Hey there,
Thank you for your answer. As instructed, I performed a clean installation after using using DDU to remove the older driver, I made sure to let the benchmark run a couple times before capturing the performance, just to make sure the game wasn't compiling any shader or something similar, but there is no apparent difference:
Hello Dean,
I took this morning to run the tests, I reinstalled the game and I also deleted the DirectX Shader Cache using disk cleanup.
Following the same methods as before, I let the benchmark run three times before starting the recordings, then I made three runs and aggregated those three runs into a single capture to get the average of the three runs. This time, I ran three more runs on the same settings but with the resolution set to 1920x1080 (16:9) and did the same process of aggregating those three runs into a single capture for the averages.
Now, about the regular gameplay performance data, it's a bit tricky to get relevant data using regular gameplay. Maps tend to perform very differently from one another, and casual/death-match servers also tend to be very different depending on the amount of players the server has. Competitive matches take way too long to measure and measuring the performance in replays (demos) is not possible because the performance you get on them does not reflect how the game runs normally. Offline with bots is also not possible due to how much more demanding it is compared to playing on online servers. Normal matchmaking casual/death-match servers take way too long to document results, because you can't choose what map you want to play, only a group of maps. As a result, you have to wait until the match is over and hope the next map is the one you desire.
So I decided to use community servers instead. I chose a server hosting the death-match mode on mirage, with 12-18 players on the server (https://xplay.gg/). I played all runs in the same server, playing normally for a bit before starting to record the performance, and I let the recording run for about 10 minutes on each run (a little bit less on Full HD runs). I found this method did produce consistent results. The results should be comparable to normal matchmaking servers, but be aware that the map being played will make a big difference in the results (I tested de_mirage).
Moving on to the results, as you can see by the graphs, reinstalling and clearing the shader cache resulted in no difference to the performance:
Running the benchmark at 1920x1080 (16:9) resulted in similar results as 1280x960 (4:3); the game runs around 22% slower on DirectX 11 compared to when running on Vulkan. The death-match recordings show a similar story as the benchmarks did.
Now, about the performance behavior specifically with driver version 101.6913, I performed a clean installation of the driver, and I didn't see much difference. The only thing that seems to stand out to me is that on the old driver, the death-matchperformance at 1920x1080 when using Vulkan is noticeably superior, but like I said, DM gameplays are inconsistent and it could just be me not seeing as much action as I did on the other run. You can also spot better averages in DirectX 11, but it's minor. TLDR performance is still way worse than Vulkan in this version.
Thanks for your time, I'll attach the tests to this post. Sorry for the boring results as they don't seem to show anything new.
Anything else just let me know.
Hey Dean,
I was a little busy, sorry for the wait, but I took today's afternoon to make the tests.
I booted into safe mode to use DDU, uninstalled the old drivers, then installed the latest version (8250), ran the benchmark 3 times before starting the capture, and again, there's no visible difference:
To me, it seems clear this is some sort of optimization issue, that requires some inside work to get it fixed, and so far, I haven't seem anything related to this issue (CS2 or DX11) mentioned in the release notes of the new drivers. CS2 is not the only DX11 or older game that has bad performance, as a matter of fact, most games that I tried that use DX11 or older have bad performance, like Far Cry 4 or GMOD (dx9). If not bad, at least worse than expected for the hardware. I just finished the remake of Silent Hill 2, and the B580 goes really well on that game, I played it on ultra settings, ray tracing on, with Xess on ultra quality most of the time, and I was getting above 60 fps in most areas of the game, except some indoor areas, but that's on the game optimization, nothing to do with the card as other cards experience the same issue. So it kind of bugs me to see the b580, which is capable of such great performance have a problem like this. The performance on Vulkan clearly shows the card can push more than that, and again, Vulkan is poorly implemented on CS2, absolutely no one but linux players use it, because it's slower than DX11 in all other cards. If my guess is correct, if the drivers were optimized for DX11 (at least for CS2) the card could push even more than the Vulkan results.
So far, no driver version has changed the outcomes, so safe to say the issue has been there for a while. Was there supposed to be something different for CS2 in this version? Or version 101.6913?
Thank you for your attention to this matter,
Gian.
