Affinity 2.0 is Hideously Slow/Buggy on Windows

[deeds]

1 hour ago, Tia Lapis said:

The problem is that Metal is way better than the available OpenCL implementations for such workloads.

 

This would only matter if they were significantly using Metal and OpenCL.

There's no indication they are significantly using these APIs and the GPU hardware. 

These benchmarks indicate most everything is being done on the CPU, and only one or two cores, at that. 

If you doubt that, examine the processes whilst doing some real world work with these apps, and you'll see the situation quite clearly. 

Open Xcode if you want to examine how little these apps use the GPU on Macs. 
 

[Nomad Raccoon]

4 minutes ago, deeds said:

This would only matter if they were significantly using Metal and OpenCL.

There's no indication they are significantly using these APIs and the GPU hardware. 

These benchmarks indicate most everything is being done on the CPU, and only one or two cores, at that. 

If you doubt that, examine the processes whilst doing some real world work with these apps, and you'll see the situation quite clearly. 

Open Xcode if you want to examine how little these apps use the GPU on Macs. 
 

Also, what happened to CUDA guys? Is it verboten? 

Why isn't CUDA a thing in Affinity? And don't tell me it's a complete rewrite of the app, cause it's not. 

 

PS:  regarding the cores, i have posted a graph showing the benchmark only using 50% of CPU power, and 60% of my GPU. And that's A BENCHMARK. Usual app usage is 5-10%

Edited December 15, 2022 by Nomad Raccoon

[slizgi]

1 minute ago, Nomad Raccoon said:

Also, what happened to CUDA guys? Is it verboten? 

Why isn't CUDA a thing in Affinity? And don't tell me it's a complete rewrite of the app, cause it's not. 

Problem with CUDA is that is only an Nvidia thing, OpenCL and HIP are multiplatform and hardware-agnostic.

[Nomad Raccoon]

1 minute ago, slizgi said:

Problem with CUDA is that is only an Nvidia thing, OpenCL and HIP are multiplatform and hardware-agnostic.

Back when CUDA was working on MacOS, it was doing pretty good, I don't know what to say https://community.adobe.com/t5/premiere-pro-discussions/shock-result-opencl-vs-cuda-vs-cpu/td-p/7354875

This seems like one of those arguments on hardware, not on performance of said hardware.

Issue at hand is we have plenty to work with, and no software to work it.

 

[Nomad Raccoon]

On 12/13/2022 at 2:56 AM, rvst said:

I'd definitely be curious to compare a Geekbench OpenCL compute benchmark with the one posted earlier in the thread. Your 3080 should blow my 3070 out of the water. 

Hard to tell what else it would be - your machine is overclocked pretty heavily as well and it's the major significant difference between your x470 chipset and my x570 chipset. The only other big difference is I have a lot more RAM, but that's more likely to slow things down, since I have all 4 banks populated (not sure what your memory config is - I see it's the same speed, but whether or not you have 2 or 4 banks populated I don't know). 

One other difference I note is I'm using Windows Pro and you're on Home, but I've never heard of any performance difference between the two. 

Hey @rvst 

My 3080 scored 201513 score , safely ranks in the 3080TI range of the Geekbench open cl benchmark. Probably because I thermally modded it myself, and it's OC-ed nicely, but geekbench does not really stress it either to be honest. (max tdp 258w, barely 70% load)

I might have some time in the weekend to properly test the CPU part of the geekbench as well, but that will require at least a Curve Optimizer run.

 

Edited December 16, 2022 by Nomad Raccoon

[rvst]

10 hours ago, Nomad Raccoon said:

So your GPU trounces mine on the Geekbench OpenCL benchmark, a full 44% faster, but is 30% slower on the Affinity benchmark.

Puzzling. 

[Nomad Raccoon]

10 minutes ago, rvst said:

So your GPU trounces mine on the Geekbench OpenCL benchmark, a full 44% faster, but is 30% slower on the Affinity benchmark.

Puzzling. 

Yes and both should be using OpenCL. Meaning the OpenCL implementation is really shoddy in Affinity. Can we get a mod/dev to explain this??

[debraspicher]

I retested my Geekbench score and it seems to vary. (Sorry, I'm not familiar with this benchmark. I usually use other programs). I edited the post above with the date. I waited to rerun, but it was closer to 200-205k this time after continuous runs.

I would love some answers as well but since @Mark Ingram is out, it may not be immediate. Also, there's still the matter of an official MSI release to see if performance is actually impacted or not by true sanded/unsanded.

[Horseflesh]

Surely there is a way for a MSIX app to have normal performance? I have gotten games from the MS store via Game Pass and they certainly didn't perform worse than .exe games. No one pixel peeps and checks benchmarks like gamers, if there was a guaranteed performance hit the game world would be out for blood. 

Anyway for fun here's my benchmark run. My machine is crummy, it's an old 6th gen i7 laptop with a mobile Nviadia 1070. 

256 / 989 / 145 / 6210 / NA / 182 / 4021 / NA

Whatever the benchmark says, perf has been fine for the kind of tasks I do, so I have no complaints yet. 

 

 

[Napkin6534]

In continuation of my previous benchmarks (see below) I add a new one with sandboxed (still damn MSIX) v203.

 

 

[Duskstalker]

before you get your pitchforks out, please consider that the affinity photo benchmark scores are easily skewed from version to version due to changed score weight or scaling. serif also stated that platform to platform is not exactly comparable.

i noticed my cpu score dropped from 2.0 to 2.03, but that doesnt automatically mean that performance dropped, it may just mean that the same performance now reports with lower scores. this is also the case with, say, cinebench for newer versions.

that said, here is my score with a ryzen 9 5900X and an intel A770 running native on win10:

 

as you can see the A770 actually scores really well in comparison to nvidia. its just amd thats so terrible. i have smooth performance with the system. if nvida runs well and intel runs well, maybe its amd and not affinity.

the truth is: amd is selling gaming cards. and amd doesnt care at all about productivity. i have had my fair share of issues with radeon cards and the intel arc cards in beta stadium actually run perfect in comparison. i had image corruption issues in capture one with openCL on a vega56 that baked into the export and it took over a year for an update (driver or program) to fix the issue - i reported that issue to amd several times. the corruption was caused by a newer version of the amdopencl.dll in the newer drivers. copying in the old radeon driver opencl dlls fixed the issue.

[SrPx]

But we are seeing people with issues that go away when disabling Open CL, while their systems actually have a nVidia card. Isn't it so ?

I realized disabling Open CL (my card is a nvidia 3060) in my preferences removes a slight (yet kind of important for good fluid work) lag I was having while painting. I told my exact settings to another forum user recently, who reported here to have quite some lag, and those solved it for that case, too. I believe it worked also with another pair of cases. So, it might not be casual or too specific.

The thing is, with it disabled and some other preferences modified, Photo 2.0.3 seems pretty fast in my side... 

And surely that is not a general rule. That is, surely the best thing is to keep Open CL on if you don't observe any issues in your workflow.

So... I'm not sure is just about AMD cards. The typical issue with AMD cards in my experience and benchmarks  (Puget Systems, etc) for graphic work is lower performance in a number of graphic apps.

With that and some other changes in preferences (and the system cured in many ways) I have to say Affinity 2.0.3 allows me now to paint very fast, fluidly. And... it has advantages in the paint system over 1.x. And one more generic improvement:  I recently discovered that... well, it might be some misconfiguration on my side, but I don't seem to be able to make the Color Studio/panel to display colors in actual Adobe RGB (even being all setup for that in AP), but -probably-  sRGB. Yet I can on 2.0.3. So there's often a very slight mismatch with Adobe RGB documents in 1.x, between the color sampled in  that panel, and the actual color painted on canvas (tested with a marquee fill, too, just in case). Not a big issue, so I keep working in some old production files  in the old version, but for new ones, I'll probably switch to 2.x. This of the color panel in a different color profile (usually sRGB) happens also  in some other apps (Krita seems to do it well, but I believe Clip Studio doesn't. And makes sense as CSP only allows you to set a color profile as preview/proof. And it works fine, so, but you have to remember to activate it with every file, and also, the issue of the mismatch with the color dialog. So, AP 1.x is not alone in that).

I was expecting that disabling Open CL would make my system unable to paint with big brushes, but it seems it yet uses the card (I don't select WARP), or, the CPU handles it well. But painting in a 4k x 4k px canvas with complex brushes (oils, etc) of 500px size is totally fluid as the best app in that. Which was not happening before (my change of GPUu might be another cause of some of that (1650 -> 3060), but maybe not all of it).

PD: As much as I like that company (AMD) I believe  that for graphic work is better to install a nvidia, even a lower rank model.

[debraspicher]

There's too many posts to link/requote buried in the thread, but if anyone is curious, @Patrick Connor posted an official unofficial but tested working version MSI installer here. I'm still seeing the same variance in GPU scores from Sandboxed and Unsandboxed, sadly... I'll wait until official though to retest on my other machines.

Sanded Official:

Unsanded Official Unofficial:

Variance: 16% ~ 18%

[Nomad Raccoon]

I honestly uninstalled the sanded version as soon as I could, but I ran tests in the unsanded one, and it looks fine. I am willing to guess that i will still have the same variance as Debra above, as that was the case with older benchmark versions as well.

One thing I want to note: the unsanded version does not have "hiccups" in using the app. Random freezes basically are present in the sandboxed app. That's the main reason I was on this thread to begin with. That and the really poor performance zooming in and out of regular images - this last part was fixed after fiddling with my own computer, but the freezes remained until I switched to this unsandboxed version.

I think the sandboxed version is just unable to use system resources in the same way as the other one, there is no other explanation for what we are seeing.

[debraspicher]

@Nomad Raccoon

Yes, I feel more confident in the MSI version.

I did start a thread in the bug section related to the sanded vs unsanded performance differences if anyone else wanted to test for a bottleneck on their machine. Many of us are using AMD CPUs, so I'm not sure if that's also a factor. Hopefully it'll assist developers to narrow down a cause. We have the solution already though which is simply to use MSI.