Radeon ROCm 1.9.1 vs. NVIDIA OpenCL Linux Plus RTX 2080 TensorFlow Benchmarks

[theriddick]

Surprised to see HBM2 fail to keep up, after all the hype about how expensive it was because it was just better, and it falls flat on its face here...

[Marc Driftmeyer]

What an absolutely moronic test. Wait until the MI60 arrives to test TensorFlow against the 2080 series.

[bridgman]

Thanks Michael. I'll see if we can get someone on our end to help...

[Michael]

In fairness, the pattern I am generally seeing here is that we lose on synthetics but compete well on real-world applications. I'll ask about the Luxmark compile errors.

Agree that it would be good to get Tensorflow etc... included in the tests. It's probably not reasonable to expect ROCm support included in NVidia's docker images, but my impression was that PTS tests usually built from source anyways.

Generally PTS does build from source, but in the instance of the NGC-TensorFlow test, it's using the NVIDIA Docker image since building it against all of the different CUDA components, etc, can be a bit finicky that it's much easier relying upon their official binary images. But will look into tensorflow-rocm and if it behaves nicely with just a Python pip install can get it added, wasn't aware the TensorFlow ROCM support was in good shape.

[bridgman]

As far as I'm concerned, it seems like the drivers just need improvement. Some of these tests show that a Vega 64 can outperform a 2080 so there's definitely real potential there. ROCm is still relatively young for such a complex project.

In fairness, the pattern I am generally seeing here is that we lose on synthetics but compete well on real-world applications. I'll ask about the Luxmark compile errors.

Agree that it would be good to get Tensorflow etc... included in the tests. It's probably not reasonable to expect ROCm support included in NVidia's docker images, but my impression was that PTS tests usually built from source anyways.

[PinkyDemon]

Would need to see how its packaged, this testing is with nvidia GPU cloud docker image.

There are some rocm dockers, but I didn't really use them. However "pip install tensorflow-rocm" works just like the tensorflow-gpu for nVidia

[Michael]

Is it possible to extend this test with tensorflow-rocm for the Radeons?
since 1.12 (I think) the rocm is part of the original tensorflow

Would need to see how its packaged, this testing is with nvidia GPU cloud docker image.

[PinkyDemon]

Is it possible to extend this test with tensorflow-rocm for the Radeons?
since 1.12 (I think) the rocm is part of the original tensorflow

[Meteorhead]

And here I thought we'll finally see the ROCm port of Tensorflow matched against NV cards.

[schmidtbag]

It is a bit shocking and sad to see AMD so far behind now in compute. Hopefully new cards soon to come will address this.

As far as I'm concerned, it seems like the drivers just need improvement. Some of these tests show that a Vega 64 can outperform a 2080 so there's definitely real potential there. ROCm is still relatively young for such a complex project.