More info on Hardware Benchmark discrepancies

This question was created in reference to: [GPU Benchmark (SynthBenchmark) gives different GPUPerfIndex values in [Content removed]

We are also encountering the issue described in the thread above. I noticed that the issue reported for the thread above has incorrect repro steps, and wanted to clarify:

https://issues.unrealengine.com/issue/UE-290402

It isn’t that you get inconsistent values on consecutive runs for within an Editor session. Rather the problem is that these values have changed significantly compared to an older version of Unreal, which throws off GPU/scalability bucketing.

For example, for a 4070, we see the following:

in UE5.4.4:

GPU Final Results:

 ... 53\.934 s/GigaPix, Confidence\=100% 'ALUHeavyNoise'

 ... 59\.555 s/GigaPix, Confidence\=100% 'TexHeavy'

 ... 56\.924 s/GigaPix, Confidence\=100% 'DepTexHeavy'

 ... 151\.189 s/GigaPix, Confidence\=100% 'FillOnly'

 ... 2\.616 s/GigaPix, Confidence\=100% 'Bandwidth'

 ... 7\.099 s/GigaVert, Confidence\=100% 'VertThroughPut1'

 ... 31\.749 s/GigaVert, Confidence\=100% 'VertThroughPut2'

GPUIndex: 658.7

in UE5.6.1:

GPU Final Results:

 ... 48\.554 s/GigaPix, Confidence\=100% 'ALUHeavyNoise'

 ... 52\.046 s/GigaPix, Confidence\=100% 'TexHeavy'

 ... 49\.008 s/GigaPix, Confidence\=100% 'DepTexHeavy'

 ... 111\.689 s/GigaPix, Confidence\=100% 'FillOnly'

 ... 2\.623 s/GigaPix, Confidence\=100% 'Bandwidth'

 ... 6\.050 s/GigaVert, Confidence\=100% 'VertThroughPut1'

 ... 30\.734 s/GigaVert, Confidence\=100% 'VertThroughPut2'

GPUIndex: 541.0

Note that several tests have discrepancies, but FillOnly stands out as the largest.

Chris

Steps to Reproduce
Run the hardware benchmark in 5.4.4, then run it again in 5.6.1 and compare the results.

Hi Chris, thanks for bringing this to our attention. I have forwarded this information to the developer team and asked them to take another look at this issue. If I hear anything new from them, I will update you immediately. Please let me know if anything is unclear.