AMD GPU Workload settings

[SoNic67]

Is there any technical explanation for the "GPU Workload" setting? I don't mean a marketing one based on the name of the setting, I mean a real explanation to what is actually happening inside GPU having that option on Compute, how it is done.

 

[Darkbreeze]

The other significant parts of 17.10.2 are the compute-oriented updates, ostensibly for cryptocurrency mining with GPUs. The new Radeon Settings "GPU Workload" toggle in the Global Settings menu switches between graphics workload optimization or compute workload optimization, although this feature is limited to select RX 500, RX 400, R9 390, R9 380, R9 290, and R9 285 cards. As for multi-GPU, Radeon Software now formally supports compute workloads for up to 12 Polaris or Vega-based RX cards on Windows 10 systems. This is perhaps in response to the number of motherboards entering the market with more than the usual number of PCIe slots.

This is LITERALLY a mining specific setting. It will either focus processing on graphic intensive tasks or mining specific compute tasks. If you are not mining, it should be set to graphics.

 

[SoNic67]

I am sorry but that is just explanation of what the name means. It's marketing to me.
That's not an explanation of HOW it is done, what is happening inside the GPU, that's what I am interested in.

 

[Darkbreeze]

Well, good luck with that, because unless you can get in touch with an AMD engineer that's willing to explain it to you, in which case most of it will likely be beyond the average persons understanding anyhow (Not that you ARE an average person, as an average person likely wouldn't be asking about this anyhow), or can obtain a white paper that specifically addresses the mining vs graphics compute workload setting, it's incredibly unlikely that any of the enthusiasts here or on any forum, aside from maybe two guys here that I'm aware of that DO have engineer level knowledge of GPU architectures but are unlikely to answer to PMs, will be able to offer that depth of explanation to you.

Basically, it is what I said it is. And any definition of purely mining related architectural compute methodology will likely offer at least minimal explanation regarding what gets focused on when a card is utilized for mining versus graphics. That function likely only optimizes an already inbuilt instruction set found in mining applications.

But I'll see if I can drag somebody in here that might know more than that.

 

[13thmonkey]

Do you really care how it does it, or what it does, the impact of it. If the latter then test it, mine for an hour get an average hash rate, change the setting and repeat.

If your actually care about the former then start digging for the white papers, and start to question how much you know about what mining is and the calculations being done.

 

[USAFRet]

Tested results:
http://www.legitreviews.com/amds-new-mining-block-chain-optimized-driver-tested_197095


"How" it works? That detail lives in someones desk drawer at AMD.

 

[InvalidError]

I am sorry but that is just explanation of what the name means.

Unfortunately, AMD hasn't detailed the technical nuts-and-bolts of what it means, so nobody knows aside from what the marketing name implies - that this setting is compute-biased in some way nobody outside AMD knows exactly how.

I'd guess that the GPU has various resources such as caches and internal buffers that are partitioned based on compute vs graphics and the 'compute' setting biases the partitioning further toward compute-centric, though I'd expect most such partitioning to be dynamic based on workload analytics these days and not need an explicit setting.

Only AMD knows. And perhaps enterprising people writing GPGPU code to poke at it and figure out what the actual differences are.

 

[USAFRet]

And I expect the actual details live behind an NDA, for people that actually know and created it.

 

[Darkbreeze]

And THAT is one of the two people I was talking about. If he doesn't know, then it's unlikely ANYBODY outside of AMD has much to say about it, at least yet, that is meaningful. Thanks to everybody who chimed in.

 

[SoNic67]

Tested results:
http://www.legitreviews.com/amds-new-mining-block-chain-optimized-driver-tested_197095
"How" it works? That detail lives in someones desk drawer at AMD.

Thanks for the link, it's a little more than nothing. Especially combined with their previous testing.
I note that that's the 17 generation driver and doesn't have "compute" optional like in the 18.

 

[13thmonkey]

So why don't you test it? , and add to sum total of human knowledge.

 

[SoNic67]

Well, I tested the two settings in PassMark. But I still don't know HOW they do it... sorry, I am curious by nature.
As you can see there is no effect on actual gaming performance, just on "compute" loads (note that the test didn't test mining per se), so personally I will leave it on "Compute", because looks like there are other things that might benefit from that.

 

[USAFRet]

Well, I tested the two settings in PassMark. But I still don't know HOW they do it... sorry, I am curious by nature.

They apply the proper amount of handwavium.

 

[13thmonkey]

I guess you need to start learning and investigating, I'm sure if you learned the right languages, learned a lot about gpu design, you could fashion programs that could test specific functionality, and then see what changes, from that you can infer what they've done.

Or

Find where amds gpu designers hang out, get them very drunk, and ask them nicely.

I suspect that compute mode is simpler than gpu mode, and maybe some functionality is bypassed, in the days of sub unit clock speeds, it may have allowed a slight clock speed increase as a result of the silicon that was deactivated.

 

[SoNic67]

That's why I put this question here - I was hoping to find some guys with more information about what actually AMD does.
Like I said, just curiosity.

 

[Darkbreeze]

handwavium.

LOL.