The AMD GPU Tweaking Thread

Hello fellow racers,

I've been fiddling a lot with Radeon Settings, Wattman, Afterburner looking to get the maximum performance out of my Vega 64 in ACC. I keep stumbling across weird issues, however. Examples:

- Poor benchmark results although clock is running way above stock settings (stock ~1530 MHz, custom ~1600 MHz)
- When launching Afterburner for monitoring purposes, Wattman settings seem to get corrupted - either right away or after system restart
- Wattman isn't resetting - it's stuck with unobvious, poorly performing settings and it doesn't reset. When I click the button, the sliders, switches and numbers switch to their default state but then return to what they were before 1 second later. Only DDU + reinstall is able to reset the settings.

So I'm fed up with that tool. I wanted to like it because I appreciate AMD's effort and I like how you can set up profiles for different games, easily. But those random changes are inacceptable and I tend to go back to Afterburner, instead.

What's your experience with this?

Now here comes the section of post 1 that I'll update with new information and also pictures of Radeon Settings and Afterburner highlighting what makes a difference, what doesn't, how many FPS can be gained with what settings etc.

Tips & Tricks for AMD GPUs

- Undervolting Vega GPUs can gain ~10 % performance in ACC. It will allow for higher clock frequencies and the GPU will usually reach them without any further tweaks to clock speed. This tweak will also reduce the GPU power consumption.
- Watch out when increasing the power limit: while during gameplay, my GPU never even reaches the stock 260 W but in the Menus it uses all of it. When increasing it to 300 W for example this power will be consumed in the menus. Together with an undervolt, this can result in system crashes. As it stands now, it seems useless to increase the power limit for ACC.
- Don't get paranoid with temperatures. There's plenty of reports regarding thermal throttling with Vegas but if you look closer, the authors usually don't provide any facts and figures. The max core temp is usually set to 85 degrees and I guess most models provide a cooling solution that can keep it in the 70s. This is perfectly fine from a hardware point of view and I haven't experienced major thermal throttling on my system. The core clock difference from a cold start (GPU running in the 50s at max) is only a few MHz higher than with a warmed up system at maximum temperature (75° for my GPU). There's pretty much no FPS difference for both these cases. I also never experienced any throttling of the HBM2 or additional stutter at high temperatures. Depending on the application, HBM2 runs hotter than the core and that's perfectly normal. The hardware documentation on HBM2 on the internet is sparse but from the sources I found the max operating temp is 125 degrees.
- In Radeon Settings, morphological filtering works for ACC. It makes no difference in FPS and blurs the image a lot.
- In Radeon Settings, setting filtering mode from "quality" to "performance" gains some ~5% of FPS in the GPU benchmark.

 

Interesting.

 

Hi,

I have a Sapphire Vega 56 Nitro+ and had some issues too. What I found out is that maxing out the clocks doesn´t change the experience at all. 1600 Mhz (1050 mV P7) or 1500 Mhz (981 mV) leads to a difference of 7 FPS max so I´ll leave it at 1500 Mhz with 981 mV on P7 in Wattman. This way I´ll get down from 220 W to 170 W usage and the fan won´t spin above 1300 rpm, temp 70°.

Unfortunately the game crashes sometimes, because the driver resets itself no matter if undervolted or not, so I guess it´s an early access problem. All other games don´t crash.

 

Thanks for joining.

7 FPS is huge, though because with your GPU that should equate to a 10 % gain.

 

I don´t think so.

Max 100 vs 107 FPS
Avg 85 vs 91 FPS
Min 62 vs 67 FPS

There is no difference gameplay wise. Maybe it´s because I use freesync.

The only difference for me is 50 W more = more heat = faster fans = more noise (at 1500 Mhz the card is almost not hearable).

 

Ah I thought you were talking about the benchmark according to the GPU benchmark thread. So with your custom settings it's a 7 % difference. I can see how it's irelevant to you in that case. I would go for noise reduction, as well.

With my custom settings I have lower FPS so it becomes a bit more relevant. It's more like 54 FPS --> 57 FPS AVG. That's 6 % which doesn't sound like much but if those 6 % keep the FPS in the FreeSync Range it's a huge difference.

If you find the time please run the GPU benchmark. We don't have a lot of Vegas in there, yet:

https://www.assettocorsa.net/forum/index.php?threads/the-acc-gpu-benchmark-thread.53426/

 

Went back to Afterburner and achieved the same results by undervolting as in Wattman.

Then as it all ran stable, I went through all of the Radeon Settings options checking which ones even do anything. Besides the Filtering set to "performance" giving a 7 % FPS gain (AVG from 67 --> 72), setting anisotropic filtering to "overwrite application settings" and the level to "2" gave another improvement by 8 % (AVG from 72 --> 78). Even on "8", I still had a 4 % improvement (75 FPS AVG). Need to check for visual differences later. So far, I didn't spot any negative impact.

In the game specific settings, that's all I could find. Besides the morphological filtering taking a toll of 1 FPS with a very questionable gain, none of the other options seem to make a difference.

I hope this is helpful to other Radeon users.

I will update post 1 later with an overview picture of Radeon Settings

 

Whilst not having a Vega series AMD GPU, it has been noted that they are memory bandwidth starved and you would achieve a much higher performance gain when overclocking the HBM2 memory. Have you tried this whilst undervolting the core clock?

 

Please provide the source where this has been noted. I don't know of any such indication. HBM2's USP is high bandwith.

@baakstaff (vega 56) and myself have overclocked our GPUs to 1000 MHz and there was no difference in FPS.

 

There have been cases like 3dmark timespy and unigine superposition where you'll see noticeable performance gains from increasing memory clock (especially superposition), but there are other applications that don't see that same scaling like 3dmark firestrike or unigine valley; it depends on the application.

That said, if you can run your memory faster, there's no reason not to since there's minimal increases in temperature and power draw (since it's at a fixed voltage) and you never see negative scaling unless you're getting memory errors.

In ACC, there's virtually no scaling going from 900 mhz memory to 1000 mhz memory with a core clock of 1570 mhz, so unless someone has a vega card with 1700 mhz+ core clock that they'd like to test, at the moment it looks to be largely core limited.

 

Aside from the gains that can be had by tweaking thermal envelopes and TDP (and apply to every game not bottlenecked by other things), when it comes to AC and some other games, the main issue is really between the game and the driver.

In certain bottleneck scenarios in VR with AC and Blackout in Blackops4, my switch from Vega64 to 1080ti provided up to 2X better framerates. Meaning that when I had dips into the 50 fps in Vega64, they were 90fps in the 1080ti. This cannot be explained by hardware alone because the 1080ti is only ~30% faster than Vega64.

But AC is too small for AMD to even look into it, and AMD has to little market share for Kunos to even look into it. So, nvidia it is, for the moment.

 

Please stop posting stuff like this here. It has nothing to do with tweaking AMD GPUs for ACC.

 

It has everything to do with it. You wont get anything close to the expected performance unless either kunos or AMD looks into it. You can tweak all you want, you can cool your vega64 with nitrogen, it wont matter. Tweaking is not the issue.

I had 9700, 9700PRO, HD4870x2, many R9 290x, Vega64. Thats 15 years with AMD altering everything from cooling, vmods, bios mods you name it I've done it.
In these cases like ACC or even AC, tweaking to gain 7fps when you are missing 40fps, will never get fixed by tweaking, ever.

 

Well, @LeDude83 is doing amazing work for all AMD users, and if i can get 67 instead of 60 frames for free I'm happy about that.
As you said Kunos and AMD aren't likely to do something about drivers, it's on our own to tweak everything.
I don't want to buy an Nvidia GPU yet (if you want to give me an rtx2080 id always take it ).
So i kindly ask you to don't blame him for the time he is investing for all of us.

 

Oh, no doubt. He is just very agressive torwards me and misses the point. If sim racing is 90% of your gaming and you insist on having AMD, its your problem. People should be aware of that.

 

Bumping this thread to pose a question: Has anyone else noticed that performance has gotten worse on AMD cards in the recent updates? I have had to turn down some graphics settings in ACC to keep framerates up.

Specifically, I used to be able to have a full grid visible without falling under 60fps (I use Chill to limit fps to 60), and now I have to stop at 16 cars or so.

 

I haven't had any issues with my RX580 / i6 6600. Stays around 60 fps on medium-high settings.

Prior to 1.0 I was getting some stutter with a full field, whether they were displayed or not, so I assumed it was CPU related. Since 1.0, I've been able to run larger fields fine.

 

Im geting worst performance with my 5700 than my gtx 1060. WTF....
In other benchmarks like heaven and userbenchmark it shows i have double the gpu performance but getting less in ACC. So cant be a bottleneck, i'm really close to 1 click away from returning and getting a 2060 super.................

 

I played 4k60 medium with vega56 pulse,
thunderstorm night race with 28 opponents without any problem or dip under 60fps
radeon settings: reset everything - apply - save
autoundervolt - apply - save
wait for vertical frame - always - apply -save
ingame frame limit 67, vsync on
temp. 69C, power 180W

 

AMD GPUs are underperforming in ACC, but not as much as you observed. See here:

https://www.assettocorsa.net/forum/index.php?threads/the-acc-gpu-benchmark-thread.53426/

Something must be wrong your end.