MP 7,1 Need your input! What benchmarks and tests would you like to see on Mac Pro with 6900XT?

[rondocap]

Thanks for that; that's awesome. 45 second improvement is nothing to sneeze at. The hard part is knowing how much of that was due specifically to the encoder, vs the rendering brute of the newer card.

What I came to find out benchmarking video type tests is that there are many variables you have to be careful with - like even SSD speed can have an impact on big files and how fast things export or render.

But these are definitely great - depends on the exact flow, but the difference certainly isn't absolutely massive from the last gen GPUs like the W5700x or Vega II - but it is a nice upgrade regardless.

 

[DFP1989]

Red 6k to H264 in Resolve, 10 min clip (H265 is similar, slightly faster)

Vega II 07:15
W6800x Duo: 06:34

That's interesting, but I guess makes sense since the stated RAW performance of the Vega II v W6800X is 14.1 TF vs. 16 TF.

The Vega II is also 25% cheaper than the W6800X on the Apple AU store at the moment which is interesting.

Edit: Wait, is your Vega II a Duo? If not that seems a bit disappointing.

 

[innerproduct]

If you would run the standard Redshift demo + benchmark on different combos of cards it would be greatly appreciated.
Especially interested on how the 6800 duo performs in comparison to a single standard 6900xt

In order to test, you would have to get a free account on Redshift3d and then follow this: https://docs.redshift3d.com/display/RSDOCS/The+redshiftBenchmark+tool

Other results can be viewed here. https://www.redshift3d.com/forums/viewthread/12843/

 

[jasonmvp]

Edit: Wait, is your Vega II a Duo? If not that seems a bit disappointing.

Remember that part of his test was a hardware export. So that specific part of the test will only use one of the two GPUs on the Duo as there's no way to use the encoders on both, simultaneously.

 

[diamond.g]

Does macOS allow you to see GPU utilization?

 

[rondocap]

At least when it comes to exporting, the difference is there - but it isn't dramatic for the most part. The Vega II is still a fairly powerful GPU for sure. The W6800x does edge it out - that test is using the encoders with H264 so something else would scale better with multiple GPU, like pro res instead.

I am going to test other performance metrics too, things like Denoise, stabilization, to see if there are bigger gaps in performance.

In things like Octane X, it's a much bigger difference that is more noticeable, but video work can be a bit ambiguous depending on the test

 

[rondocap]

Does macOS allow you to see GPU utilization?

It does, using iStat app - let's you know what is being used or activity monitor has some info as well

 

[jasonmvp]

At least when it comes to exporting, the difference is there - but it isn't dramatic for the most part. The Vega II is still a fairly powerful GPU for sure. The W6800x does edge it out - that test is using the encoders with H264 so something else would scale better with multiple GPU, like pro res instead.

I don't like the way Resolve calls the process of creating a new video file: "rendering". It's a half-misnomer, if you will, and leads to potentially confusing conversations.

Rendering - the process of compositing each video frame, including any effects (color, etc) required.
Exporting - the process of creating an output file.

During the first of those, multiple GPUs can be used. And that's where @rondocap's new card will really shine is in the actual rendering part of the process. The second of those is where the encoding happens. Asking the system to do hardware encoding (eg: h.264, h.265, et al) will only use one GPU.

 

[rondocap]

I don't like the way Resolve calls the process of creating a new video file: "rendering". It's a half-misnomer, if you will, and leads to potentially confusing conversations.

Rendering - the process of compositing each video frame, including any effects (color, etc) required.
Exporting - the process of creating an output file.

During the first of those, multiple GPUs can be used. And that's where @rondocap's new card will really shine is in the actual rendering part of the process. The second of those is where the encoding happens. Asking the system to do hardware encoding (eg: h.264, h.265, et al) will only use one GPU.

Yeah exactly - if you export in pro res instead then it does utilize multiple gpus again, along with the cpu/afterburner too.

 

[rondocap]

I don't like the way Resolve calls the process of creating a new video file: "rendering". It's a half-misnomer, if you will, and leads to potentially confusing conversations.

Rendering - the process of compositing each video frame, including any effects (color, etc) required.
Exporting - the process of creating an output file.

During the first of those, multiple GPUs can be used. And that's where @rondocap's new card will really shine is in the actual rendering part of the process. The second of those is where the encoding happens. Asking the system to do hardware encoding (eg: h.264, h.265, et al) will only use one GPU.

Jason, I saw on your link that you are also a caselabs owner, me too! I have the aquaero as well. Small world!

 

[diamond.g]

It does, using iStat app - let's you know what is being used or activity monitor has some info as well

Cool, see the encoder (ASIC) question is tricky because as far as I can tell AMD only adds new codecs to them, not improve performance for each version. If you see a performance improvement it is likely because the GPU has "fallen back" to a hybrid model where it has to use the compute to do stuff that the ASIC cannot do. Which would be faster on the RDNA line (yay high clocks).

 

[randy85]

There are just so many variables with hardware utilisation on Resolve exports. I just exported 4k prores to 9:16 h264 in HD/4K.

The HD used 2 GPUs and Afterburner and exported a 14 min sequence in 2m40s, whereas the 4K mostly used the CPU was much slower at 35m10s. I'm guessing it's something to do with 9:16 not a standard resolution, but goes to show even if you trick your system out, it may not always be utilised by the software.