Drivers & Software

[ketxxx]

The title says it all. I'm currently gathering data for a very in-depth GPU article including driver performance and I've been absolutely dumbfounded by the horrendous single and multi-threaded driver performance in the 3DMark API tests. The drivers I've looked at so far for the AMD side are Adrenaline 18.10.2 and Adrenaline 19.1.1, I of course did not take the API test as fact but upon running a plethora of DX11 benchmarks have found the API test results to also represent actual game performance in several scenarios the AMD card (RX580) being used as part of the testing is being completely trounced by an nvidia equivalent GPU (GTX980), for example in Metro Last Light with SSAA enabled the AMD card scores a rather paltry 48.1FPS average while the nvidia card clocked on the GPU to match the RX580 in frequency (1.35GHz) scores 56.3FPS average. Without divulging too many details about the article where its a work in progress I'll say that surely the AMD card should at least be equalling the 980 all of my tests all point to one reason why the RX580 is doing so badly - terrible single and multi-threaded driver level DX11 optimisation.

Any chance of getting someone from AMD to get in touch with me to discuss these many findings I have come across in the process of getting this article together? In 5 out of 6 DX11 tests performed the AMD card gets the pants beat off it these performance problems are far too widespread to be isolated instances and without giving away numbers the AMD card in the 3DMark API test is more than 100% SLOWER than the nvidia card in the multi-threaded portion of the DX11 test.

[pokester]

First, nobody from the AMD development team is here. This is a user to user forum only it says this right in the forum headers so AMD won't be answering your request. I suggest you contact them through the ways available on their contact page.

The GTX 980 is a faster card overall than the RX 580. The 980 is also a prior generation flagship card and the 580 is a current generation mid level card and it's Nvidia counterpart is the 1060, the RX 580 is pretty neck and neck with that card at 1080 and slightly outperforms it at 1440p.

The following is just one example showing this, but this example does not reflect your suggested results. I can't begin to figure out how you get what you get when you really have explained nothing.

https://gpu.userbenchmark.com/Compare/Nvidia-GTX-980-vs-AMD-RX-580/2576vs3923

[ketxxx]

Keeping it simple, the GTX1060 is largely considered GTX980 performance at the mid level, thus the point that is easy to see for anyone knowing this is that the RX580s DX11 performance in comparison is way, way below par and should be a lot closer to the 980 in DX11 than it is.. As for nobody from AMD frequenting their own community, well thats rather odd I remember when people from AMD would frequent Xtremesystems, Guru3D, etc. I'd wager someone looks through the AMD community to collect anything of interest but thats it - I used to have to do that exact thing when working for Mushkin Memory. Regarding specifications of each card, while you certainly can't compare them apples to apples as the architecture of each card is wildly different you can math out some crude numbers and extrapolate but this is all getting besides the point, which is an apparent severe lack of driver level optimisation bottlenecking the GPU. Without writing an essay that is the long and short of it.

[pokester]

You are certainly welcome to your own opinion and benchmark results. They however do not match any mainstream information from all the normal avenues. As an RX 580 and a 1060 6gb owner I can tell you my experience is far from yours. Good luck in your testing though.

[waltc]

He's just another nVidia FUDster...;)  The account is brand-new as of the 18th.  Yawn. 

[hitbm47]

Hi @ketxxx  yes DX11 draw call performance on AMD is still awefull.

Firstly, it does more draw calls single threaded than multithreaded in 3DMark, and secondly I get about 1.5x the draw calls single threaded with GTX 1060 and 2x more than my RX 480 in Multi-threaded draw call performance.

For example, on my FX 8350 and RX 480:

~ 1 050 000 (MT),   - 1 070 000 (ST)

GTX 1060

- 2 120 000 (MT),   - 1 640 000 (ST)

And even though DX12 and Vulkan can do 20x more on my system as well, most games don't implement them that way and then you get even slightly worse performance than DX11; for example, Serious Sam 4 and Dues Ex Mankind Divided.

Test with something like FX 8350 / i7 870 with a Radeon card and you will experience the unacceptable draw call performance. Furthermore, before someone says CPU bottleneck, this doesn't even max out any single core on one of those CPUs I mentioned, really don't know what AMD did with their DX11 instruction submission.

[Ray_AMD]

Since you have already opened threads on this topic, there is no need to resurrect this thread from 2019.

[joecrow]

Hmmn

I am no expert but your findings seem to be contrary to several other reviews which indicate the RX 580 performs very favorably against for example the GTX 1060 with both DX11 and more particularly DX12 gaming. Are your results significantly different with much older drivers? Is you comparison with the GTX 980/DX11 valid?

[ketxxx]

Results are valid, the reason I ended up testing the Adrenaline 18.10.2 drivers which are a fair bit older than 19.1.1 is because of this terrible DX11 performance. The games used for the DX11 portion of the performance testing the RX580 does the worst in are Metro Last Light (-14% with SSAO, -11% without SSAO), Alien Isolation (-14%), Shadow of the Tomb Raider (-10%, DX11 mode). Those figures are based on average and not the minimum frame rates.

[noodles59]

Terrible performance of just ~10% less sounds like an oxymoron to me and why talking about vintage DX11 when DX12/Vulkan is 20x faster than DX11 (on AMD)?