Affinity 2.0 is Hideously Slow/Buggy on Windows

[debraspicher]

4 hours ago, Nomad Raccoon said:

Hi,

 

I can jump on the RTX 3080 / 5950X bandwagon here, the V2 is atrociously slow compared to V1, I can't even zoom in on a 2k image without the software lagging.

I can already tell you what the problem is, but you won't like it.

It's 4 little letters - MSIX

Microsoft store apps by design work subpar in terms of performance inside Windows desktops - it's crazy , I know, right?

I am looking for an x86/x64 msi installer at this point for V2, that would immediately solve half the problem in terms of performance.

The other half I think is down to the really poor optimization of the threading, I can see the process hogging on 1-2 cores at most when I have 32 available.

 

That would be hilarious if that's what ends up mattering.

[Nomad Raccoon]

1 minute ago, debraspicher said:

That would be hilarious if that's ends up mattering.

I am pretty much 95% sure at this point, since I was getting WHEA 117 kernel errors in the benchmark with HDR on. 

I haven't seen those in a long time. 

Anyway, my biggest confusion is the combined scores, they are all over the place, some have way older hardware and better combined scores, to me that yells bad optimization somewhere?

[Napkin6534]

31 minutes ago, Nomad Raccoon said:

really poor optimization of the threading, I can see the process hogging on 1-2 cores at most when I have 32 available.

That really grinds my gear everytime I see this. Why do we have all those cores when programmers do not even use them? Such a waste.

[Napkin6534]

I have an older version still on disk. Tried to run the benchmark on both versions. I attach the result.

 

EDIT: I ran unsandboxed benchmark, and the results are attached below,

 

EDIT2: I cannot add a reply, there is some silly limit on post number, so I have to edit this post again.

Below you can find the results of a "sandboxed" benchmark on another computer with Ryzen 3100 + 32 GB of 2133 MHz memory + nVidia 1080ti

Edited December 13, 2022 by Napkin6534
additional, fourth screenshot

[Nomad Raccoon]

21 minutes ago, Napkin6534 said:

I have an older version still on disk. Tried to run the benchmark on both versions. I attach the result.

 

Yeah okay, that pretty much confirms what everyone was expecting, GPU is not used at all in Affinity atm lmao.

[rvst]

59 minutes ago, Nomad Raccoon said:

my guess is my results are better than other RTX 3080 or 5900/5950x

Your results are not better. I have a 3070 and I got 30% better performance than your 3080 on the test results I reported above. 

On 12/11/2022 at 4:21 AM, rvst said:

476/4782/982/13688/-/1266/7206/-

That said, I ran that benchmark above on an unsandboxed version of Affinity 2.

When running in the sandbox, the GPU performance indeed takes a big hit.

454/4667/950/10858/-/1184/6263/-

Those two benchmarks are directly comparable since they're the same version on the same machine, just sandboxed vs unsandboxed

I knew there was a performance hit for UWP apps, but not of that magnitude. 

Easy to test yourself - just use 7zip to unzip the MSIX installer and run the photo.exe directly from the unpacked folder. 

[debraspicher]

That's a deal breaker for MSIX, imo.

[debraspicher]

Well.. only one run each, but wow. Summary: GPU Single/Multi (Sandboxed vs Unsandboxed) 10640 vs 12809 ... a bit less than 20%.

@Patrick Connor You might want to look into this?

Strangely enough the "combined" GPU score stays more or less the same. Though I'll be honest, I really don't put that much stock into the benchmark itself and what it claims to represent.

Sandboxed:

Quote

Unsandboxed:

Quote

 

[Nomad Raccoon]

40 minutes ago, rvst said:

Your results are not better. I have a 3070 and I got 30% better performance than your 3080 on the test results I reported above. 

That said, I ran that benchmark above on an unsandboxed version of Affinity 2.

When running in the sandbox, the GPU performance indeed takes a big hit.

454/4667/950/10858/-/1184/6263/-

Those two benchmarks are directly comparable since they're the same version on the same machine, just sandboxed vs unsandboxed

I knew there was a performance hit for UWP apps, but not of that magnitude. 

Easy to test yourself - just use 7zip to unzip the MSIX installer and run the photo.exe directly from the unpacked folder. 

UWP apps are horrible, and my hunch was correct regarding msix, thanks for the tip for how to unsandbox them.

 

 

 

[debraspicher]

40 minutes ago, rvst said:

Your results are not better. I have a 3070 and I got 30% better performance than your 3080 on the test results I reported above.

Husband has a 3070, so I'll have him install the program and run the test later, albeit his CPU is not as good as yours (5600X).

[Patrick Connor]

1 hour ago, rvst said:

Easy to test yourself - just use 7zip to unzip the MSIX installer and run the photo.exe directly from the unpacked folder

FYI This is unsupported and will likely stop working in a future release (despite the results above). Patching will not work, nor will a number of other application features.

26 minutes ago, debraspicher said:

You might want to look into this?

Hmmm, Interesting. I am sure Mark Ingram might have been surprised by those results. Not clear whether that difference is going to be generically true though, but as you say some similar combined numbers

[rvst]

1 hour ago, Patrick Connor said:

FYI This is unsupported and will likely stop working in a future release (despite the results above). Patching will not work, nor will a number of other application features.

Yes we know. This was just to test the performance difference between a sandboxed and unsandboxed variant, not an exhortation to other forum users to run Affinity in this way. 

1 hour ago, Patrick Connor said:

Not clear whether that difference is going to be generically true though, but as you say some similar combined numbers

I suspect it is generic, since two of us just got the same results - we both see a massive hit

[Tia Lapis]

Strangely Blender is working perfectly fine in the MS Store version...

[Nomad Raccoon]

 

Did 2 quick tests - left is UNSANDBOXED ---------------------------------------------------------- SANDBOXED

Both results worse than earlier but the difference is clear on GPU, even on combined. Sandboxing hurts the performance. And tbh overall something is clearly wrong as 3080 should be doing twice those raster numbers right? I think there was an example unsandboxed with around 20k for raster - simple GPU.

I still believe UWP is only part of the problem, as I initially suspected. 

[debraspicher]

Tested our living room machine and my hub's machine. Yup.

EDIT: FWIW, the 5600x machine is running other processes. Husband isn't home so the only thing I closed was browser, no idea what else he had open.

5600x Sandboxed

GPU Score 9963 vs 11473 (14-15%)

Quote

Unsandboxed

Quote

Ryzen 5 3600 Living Room Sandboxed

GPU Score 7868 vs 9130 (14%)

Quote

Unsandboxed

Quote

 

[debraspicher]

20 minutes ago, Nomad Raccoon said:

Did 2 quick tests - left is UNSANDBOXED ---------------------------------------------------------- SANDBOXED

Both results worse than earlier but the difference is clear on GPU, even on combined. Sandboxing hurts the performance. And tbh overall something is clearly wrong as 3080 should be doing twice those raster numbers right? I think there was an example unsandboxed with around 20k for raster - simple GPU.

I still believe UWP is only part of the problem, as I initially suspected. 

Hard to say, because the benchmark version is different, they could've altered the way the GPU sample is taken.

[Nomad Raccoon]

Yeah I've restarted and added my normal settings, results remain similar : Multi CPU is lower in unsandboxed, but every other result is higher, by up to 20% higher in the case of GPU raster.

I see the same trend in Vector Multi CPU results  for @debraspicher  . What I don't understand is how someone with a 3070 reach a higher score on this benchmark version than we do with our 3080.

 

And more interestingly how did someone reach 22372 score  with a 3080 in 11021 Benchmark version and we can only reach 10-12k in version 20000.

[Nomad Raccoon]

1 minute ago, debraspicher said:

Hard to say, because the benchmark version is different, they could've altered the way the GPU sample is taken.

Well that's easy to verify, just find someone with a different GPU and run an older benchmark, but I doubt ONLY that one test would have its score halved, it would impact the Combined GPU scores too, wouldn't it?

[debraspicher]

Just now, Nomad Raccoon said:

Yeah I've restarted and added my normal settings, results remain similar : Multi CPU is lower in unsandboxed, but every other result is higher, by up to 20% higher in the case of GPU raster.

I see the same trend in Vector Multi CPU results  for @debraspicher  . What I don't understand is how someone with a 3070 reach a higher score on this benchmark version than we do with our 3080.

There was a 3060TI narrowly not beating a 3090 in the V1 benchmark charts lol. @MikeTO  (I believe) kept an extensive chart, but can't access the thread now. It seems correlated to CPU pairings, other factors. And tbf, I do have quite a bit running in the background that I could still clear away and my rig is not optimized to run benchmarks.

[Nomad Raccoon]

7 minutes ago, debraspicher said:

There was a 3060TI narrowly not beating a 3090 in the V1 benchmark charts lol. @MikeTO  (I believe) kept an extensive chart, but can't access the thread now. It seems correlated to CPU pairings, other factors. And tbf, I do have quite a bit running in the background that I could still clear away and my rig is not optimized to run benchmarks.

Mine's pretty damn clean and running pretty high frequencies on everything. It would ace any other benchmark you can throw at it. There is a lot of variance in the results first of all, I can repeat the test and get pretty wild values from one run to another, even they mention we should average them. 

 

But that's not my problem here, it's the fact that the software runs SLOW, and gpu-z confirms that the GPU is not being used, while process explorer is telling me that both GPU and CPU are not used at all. 

It barely reaches a 70% peak on CPU during the benchmark. The test is over before it even starts as far as I see. Same for GPU. Maybe MacOS handles this faster, but besides that I see absolutely no reason why these tests would mean anything at this point.

 

[rvst]

52 minutes ago, Nomad Raccoon said:

 What I don't understand is how someone with a 3070 reach a higher score on this benchmark version than we do with our 3080.

My rig is not stock configuration, so you should take that result with a pinch of salt. 

It's overclocked and heavily optimised. It can do ~5.1ghz on the best core, although I throttle the maximum CPU temperature to 80 celsius, so practically speaking it tops out at about 4975Mhz as higher tends to push it over 80 celsius. It can reach an all-core clock of close to 4800Mhz.

If you look at the Geekbench results I posted earlier in the thread and compare them to others' results on the Geekbench website, you'll see that it's on a par with overclocked systems running at 5Ghz. 

The 3070 is a Gigabyte Auros Master - Its core clock is 1845Mhz vs the reference card of 1725Mhz, so a bit faster than the reference card.

That said, it's running stock, so I have not attempted to OC it. It's not faster than anyone else's 3070 on the Geekbench OpenCL compute benchmark, so the performance difference is not down to the GPU, it's probably due to my system - your 3080 benches much faster than my 3070 in Geekbench. I'd be interested to see what your compute benchmark is if you run Geekbench on your system. 

The key difference is that I have a more capable motherboard, an ASUS ROG Crosshair VIII Dark Hero (x570). All PCIe lanes, unlike yours and Debra's are PCIe 4.0. So it's going to have generally better performance when throwing around large amounts of data over the PCI bus. The RAM is running at 3600MHz and I have Resizeable BAR configured, although it didn't actually make any noticeable difference in performance with Affinity benchmarks, as I tested it on and off. 

[rvst]

57 minutes ago, Nomad Raccoon said:

And more interestingly how did someone reach 22372 score  with a 3080 in 11021 Benchmark version and we can only reach 10-12k in version 20000.

Affinity v1 benches about 2x higher on the single GPU raster test. It's because of the difference in versions. The results are not directly comparable with v2.

Here's a v1 benchmark on my system - almost 20k

 

[debraspicher]

40 minutes ago, rvst said:

The 3070 is a Gigabyte Auros Master - Its core clock is 1845Mhz vs the reference card of 1725Mhz, so a bit faster than the reference card.

I could test an OC to see if it would make a difference in benchies. I have the headroom, I just haven't felt the incentive. These are the specs on what I have running. By core clock, I'm guessing you mean boost clock (Mine is 1755).

I really don't think it's the difference in PCIE bus, personally, but it's certainly possible your motherboard is more optimized on a newer chipset. At least, I don't see signs the Affinity app to be testing it that aggressively. I'd imagine we'd be seeing more in GPU-Z and in temperatures... we're also assuming it's optimized.

As for my machine, I'm running an all core -30 undervolt (as it says sig) @ 5ghz with no WHEA errors for at least a few months. I'm sure I could optimize to higher on my best cores if I work with cores individually, but I don't have the time or patience to do the stability testing necessary to make the most of it when I have an almost toddler running on the loose daily in the background. I also still need this machine active for work.

I'm on air cooling as well which is a major factor for the 5800X and higher core density. It will hit the thermal wall faster than the other processors. I don't see how I could lower the thermal limit and still be capable of the same performance without water cooling.

Edit: All core in OCCT is 4850Mhz

[Nomad Raccoon]

38 minutes ago, rvst said:

My rig is not stock configuration, so you should take that result with a pinch of salt. 

It's overclocked and heavily optimised. It can do ~5.1ghz on the best core, although I throttle the maximum CPU temperature to 80 celsius, so practically speaking it tops out at about 4975Mhz as higher tends to push it over 80 celsius. It can reach an all-core clock of close to 4800Mhz.

If you look at the Geekbench results I posted earlier in the thread and compare them to others' results on the Geekbench website, you'll see that it's on a par with overclocked systems running at 5Ghz. 

The 3070 is a Gigabyte Auros Master - Its core clock is 1845Mhz vs the reference card of 1725Mhz, so a bit faster than the reference card.

That said, it's running stock, so I have not attempted to OC it. It's not faster than anyone else's 3070 on the Geekbench OpenCL compute benchmark, so the performance difference is not down to the GPU, it's probably due to my system - your 3080 benches much faster than my 3070 in Geekbench. I'd be interested to see what your compute benchmark is if you run Geekbench on your system. 

The key difference is that I have a more capable motherboard, an ASUS ROG Crosshair VIII Dark Hero (x570). All PCIe lanes, unlike yours and Debra's are PCIe 4.0. So it's going to have generally better performance when throwing around large amounts of data over the PCI bus. The RAM is running at 3600MHz and I have Resizeable BAR configured, although it didn't actually make any noticeable difference in performance with Affinity benchmarks, as I tested it on and off. 

In fact, I have a X570 Aorus Master motherboard, it’s all pcie 4.0, even the the ssds. 

I didn’t OC my Cpu that hard- i think it’s default PBO atm,  then again 5950x doesn’t OC thaaat much tbh. I will check if going around 5ghz single core changes things but it’s still a silly test if you look at the cpu/gpu usage of the test. The variance in results also supports this.

 

 

 

 

 

 

[rvst]

8 minutes ago, debraspicher said:

I could test an OC to see if it would make a difference in benchies. 

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. 

9 minutes ago, debraspicher said:

I really don't think it's the difference in PCIE bus, personally

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.