PyTorch 1.8 Released With AMD ROCm Binaries

[haydn-j]

You can use tensorboard with pytorch too.

Yep! And there are fantastic libraries like pytorch-lightning.

[qarium]

AFAIK, he lacks a MI100 and that's the only AMD GPU with Matrix Cores. Otherwise, your performance is going to be potentially quite a bit slower than any Nvidia GPU with Tensor cores.
I mean, you could always just use fp32, but then it's no longer a very realistic benchmark (other than showing things like relative driver overhead of each stack).

just wait for the second generation of CDNA with chiplet design. i am 100% sure they will use a single chiplet to also produce a cheap variant of it. then you can buy one for like 700€..

[boffo]

So, let me get this straight, boffo and Steffo both registered their accounts in 2013 and both have nearly the same # of posts? Interesting...

LOL!!! You got the OCD eye!

[Steffo]

So, let me get this straight, boffo and Steffo both registered their accounts in 2013 and both have nearly the same # of posts? Interesting...

To be honest: Your kind of your perception is a little creepy. I don't think, you do yourself a favour with selective perception which has basically no point. I don't even have an AMD graphic card for reasons like power consumption and bad support for neural networks. - The latter is changing now, the former hasn't changed much. NVIDIA cards are still more efficient than AMD cards.

[oleid]

The AMD support is really huge! I hope, this will be a game changer in the AI market!

Well, it is not as if tensorflow support is not available for some time now.
You can get tensorflow-rocm via pip.

[Steffo]

Well, it is not as if tensorflow support is not available for some time now.
You can get tensorflow-rocm via pip.

This is a fork and not officially supported.

[oleid]

This is a fork and not officially supported.

True, but it works fine.

EDIT: I don't want to downplay the value of getting things mainline. I just don't think it's a "game changer." Maybe for torch, because that wasn't so easy to use with ROCm before. But definitely not for tensorflow.

[bridgman]

AFAIK tensorflow support for ROCm has been upstream for over a year now, with regular builds off the upstream source code.

https://github.com/tensorflow/tensorflow

The "tensorflow-rocm" repo is now called "tensorflow-upstream" and used for ongoing development. We do build "tensorflow-rocm" docker images, however.

Pytorch upstream support is relatively new though.

[coder]

LOL!!! You got the OCD eye!

That's not even all. I further computed that your average post interval is 8.75 days, while Steffo's is 9.05 days. I wondered how close the ~7 months between your account registrations came to compensating for your different post counts.

[coder]

To be honest: Your kind of your perception is a little creepy. I don't think, you do yourself a favour with selective perception which has basically no point.

If we're being honest, it really was a funny coincidence to have two similar-sounding accounts post similar sentiments, right at the start of the thread, and with registration dates and activity level that are further closely aligned. I was mostly having a laugh about it, since it's almost too inconsequential to be nefarious.

I don't even have an AMD graphic card for reasons like power consumption and bad support for neural networks. - The latter is changing now, the former hasn't changed much. NVIDIA cards are still more efficient than AMD cards.

RDNA and RDNA2 have both made a lot of progress on power-efficiency, at least for gaming. Meanwhile, Nvidia has been delivering sub-par efficiency improvements since Pascal, which have been masked by their addition of Tensor Cores. AMD introduced Matrix Cores in their M100, but we can only hope and wait for them to trickle down to more affordable GPUs.

At work, I've had to use Team Green for the reasons you cite. We jumped on that bandwagon back in the Pascal era and haven't yet had cause to reconsider, though we're also dabbling with OpenVINO (Intel).