i9-9900k & RTX 2080 vs. Ryzen 7 9800X3D & RTX 5070Ti

[Tauri141]

I have two PCs with the configurations as in the title.
I run this code on both PCs whisper-faster.exe video.mkv --language Russian --model large-v2 --compute_type float16 --output_dir source
As a result, on the PC with Intel, subtitles are ready in 7 minutes, and on the PC with AMD in 12 minutes.
CUDA version on the first PC 12.7, on the second (newer) 12.8
How can it be that on more older, more weaker hardware, it is done 30% faster?

[Purfview]

whisper-faster.exe

That must be something very outdated.

How can it be that on more older, more weaker hardware, it is done 30% faster?

Test latest Faster-Whisper-XXL!
Btw, if you do speed benchmarks then use -fallback None option.
Ensure that all GPU hyper-parameters are same. Like 'Hardware Accelerated GPU Scheduling'. [this option alone can give 30% difference]
Check if NVIDIA driver is same version.

[Tauri141]

Installed Faster-Whisper-XXL_r245.2_windows.7z
On Intel in 5 minutes 36 seconds (video driver version 566.36)
Transcription speed: 8.61 audio seconds/s
On AMD in 11 minutes 35 seconds (video driver version 572.83) Transcription speed: 3.99 audio seconds/s

Hardware Accelerated GPU Scheduling on Intel is disabled. On AMD is enabled.

The PC on Intel is a working PC, I can't update the video driver without permission from management.

[Purfview]

Then you can disable that option on AMD PC and you can roll back driver to 566.36.

Sometimes [pretty often] "newer, faster & better" is just a marketing stun. 😊
Btw, check if outputted subtitles are same/similar.

[Tauri141]

I disabled Hardware Accelerated GPU Scheduling on AMD.
Time 11 minutes 30 seconds. Disabling this parameter did not help.
I will not roll back the driver from 572.83 to 566.36 (Drivers 566.36 do not know what the RTX 50XX series of video cards is).
The output subtitles are similar.

[Purfview]

With different drivers you don't know what you are bechmarking, maybe difference comes from the drivers.

[Tauri141]

The work PC is old and the drivers are old, but the program works faster than on everything new... somehow it's not right.
For me, the fact that it works faster on the work PC is a plus. I just wanted to compare its work on a personal PC for fun.
In any case, it turns out to be a downgrade for programm, since it works worse on more powerful hardware and newer software. Or it's not a downgrade of the program, but of the software from NVIDIA.

[Purfview]

Rant about "more powerful hardware":
Turbo Boost v3.0 is a scam and in some tasks a 10(!) years older equivalent CPU performs actually faster, when all tech specs tell you that it should be more than twice slower. (Yeah, FU Intel!)

[Purfview]

Btw, try to benchmark batched mode, maybe there it wouldn't be that slow: --batched

[Tauri141]

With the --batched parameter.
On AMD 2 minutes 45 seconds.
On Intel 2 minutes.
But the subtitles created in 400 symbols per line!

[Purfview]

But the subtitles created in 400 symbols per line!

Use --sentence or --standard for long lines to be split.

[Tauri141]

Now everything is super.
Although it is not clear why on the older assembly it is done for 2 minutes, and on the new one 2 minutes and 45 seconds.

[Purfview]

Btw, you can try --batch_size 16 or --batch_size 32 [default is 8]

[Tauri141]

With parameter --batch_size 16 on AMD 1 minutes 53 seconds, on Intel 4 minutes 12 seconds (It's strange, by default, you say 8, the speed is higher)

With parameter --batch_size 32 on AMD 1 minutes 28 seconds.
On Intel error (Probably 8 GB of video memory is not enough)

[Purfview]

So, you found where AMD/5070Ti is faster.

It's strange.

Why it's strange? Do you think that increasing batch_size the speed should infinitely increase too? :)
It's slower because that batch size is sub-optimal for that system.

You can try various batch sizes to find the optimal one.

[Purfview]

Btw, few minutes test probably is too short to have accurate benchmarks, in the light that CPU makers engage in scummy techniques to skew such short benchmarks.

Get a file where a test would run at least for 30 minutes. [you can join same file multiple times]

[Tauri141]

I conducted all these tests on the same video, 45 minutes long.
Conclusion, as a result, I got a high speed increase thanks to the --batched command, for me this is enough.

[Purfview]

I meant the actual test duration should be longer for accuracy.