Vega 12/20 Added To AMDGPU LLVM, Confirms New GCN Deep Learning Instructions For Vega 20

[coder]

To my knowledge, in currently fails. TensorFlow is currently nVidia or cpu only. Maybe it will happen some day, but I've heard that so often, I'll only believe it when I see it.

Sure, it has open issues, but I think it's simplistic to say it fails.

[AndyChow]

Sure, it has open issues, but I think it's simplistic to say it fails.

It fails to build, is what I mean. Current git demands cudnn-7.1.2-1, and trying to trick it simply fails, so far. But I haven't tried that hard. Even the current page admits it fails to build, but I thought I'd try my 20-minutes "pass or smash" skills. No go so far.

[olesalscheider]

BTW, cpu-only isn't that bad. Most models can be trained in a few hours, and most models don't beat a carefully built probabilistic model, on average. Just throwing more computing power isn't a good solution. Many times these young guys come in and build something they think is great, I build a stupid gamma model, or a two variable beta, and my predictions beat theirs, hands down. Or, and this is humiliating, an exponential or Poisson. So it's nice technology, but removing careful thought and just letting the model define itself through more data and more computing power is IMO not that productive. We aren't there yet.

Well, that might be true for some models. And of course it does not make sense to throw away prior knowledge if you have it.
But there are also models where you can't really model everything. Take image processing for example. Even here, the models can be greatly simplified by using a bit of prior knowledge (e. g. that objects in an image are usually translation invariant to some degree -> you can use a CNN instead of a fully-connected NN and you have to learn a lot less parameters). And still, these models need days to weeks on ten(s) of high performance GPUs...

[GunpowaderGuy]

For anyone wanting cross vendor gpu acceleration , I recommend using keras through the plaidml backend

[GunpowaderGuy]

Michael should talk about Intel's neural compute stick

[bridgman]

It fails to build, is what I mean. Current git demands cudnn-7.1.2-1, and trying to trick it simply fails, so far. But I haven't tried that hard. Even the current page admits it fails to build, but I thought I'd try my 20-minutes "pass or smash" skills. No go so far.

When you say "current git" do you mean the repo that coder linked to (which builds and runs on ROCm AFAIK) or latest Tensorflow git ?

I wouldn't be surprised if latest Tensorflow git failed because of some new dependencies, but AFAIK we are running Tensorflow and other frameworks today.

[Tomin]

I wouldn't be surprised if latest Tensorflow git failed because of some new dependencies, but AFAIK we are running Tensorflow and other frameworks today.

This reminds me. I have one Phenom II system so there is no PCIe Atomics there. Is there any low(ish) power GPU that can be run with ROCm on that system? RX {460,550,560} are not an option because ROCm needs the atomics with Polaris so that wouldn't work. I'm looking for something that has low idle power and not too great demands on full power (just one or no PCIe power connector for example). The card may be second hand. Most GCN cards that don't need the atomics seem to be quite power hungry. If there is no such card then the time to move on from Phenom has probably come. (Or maybe I should get one of those Tensor USB sticks...)

[coder]

RX {460,550,560} are not an option because ROCm needs the atomics with Polaris so that wouldn't work.

I think this requirement is gone, in the latest AMD KFD. Is that true of the mainline 4.17 kernel? Anyway, I didn't think the PCIe atomics requirement was specific to the generation of GPU.

I'm looking for something that has low idle power and not too great demands on full power (just one or no PCIe power connector for example).

Depending on the above, your best bet would be a 4 GB RX 560, IMO. RX 460 wouldn't be a bad compromise. Either way, pay attention to the number of shaders (ranges from 896 to 1024). Or even RX 550 (now available in both 512 and 640 shader versions...)

Otherwise, maybe get an old Bonaire card? You can find the various models and specs, here:

https://en.wikipedia.org/wiki/List_o..._RX_400_Series

(Or maybe I should get one of those Tensor USB sticks...)

Those are toys, IMO.

If you're prototyping an embedded solution and need a way to test/evaluate the inferencing performance of that chip, or if you need to add some inferencing horsepower to a Raspberry Pi class machine, then go for it. Otherwise, any real GPU should kick them to the curb.

[Tomin]

I think this requirement is gone, in the latest AMD KFD. Is that true of the mainline 4.17 kernel? Anyway, I didn't think the PCIe atomics requirement was specific to the generation of GPU.

I was reading this: https://github.com/RadeonOpenCompute...supported-cpus

Actually they say that

We do not support ROCm with PCIe Gen 2 enabled CPUs such as the AMD Opteron, Phenom, Phenom II, Athlon, Athlon X2, Athlon II and Older Intel Xeon and Intel Core Architecture and Pentium CPUs.

But does that mean they don't provide support or that it will not work?

Elsewhere they give even shorter list of supported hardware:

https://rocm.github.io/hardware.html

Anyway, they don't really advertise that it would work on something that doesn't support PCIe atomics apart from this:

Experimental support for our GFX7 GPUs Radeon R9 290, R9 390, AMD FirePro S9150, S9170 note they do not support or take advantage of PCIe Atomics.

Which in my case is not helpful since those are pretty power hungry GPUs.

Depending on the above, your best bet would be a 4 GB RX 560, IMO. RX 460 wouldn't be a bad compromise. Either way, pay attention to the number of shaders (ranges from 896 to 1024). Or even RX 550 (now available in both 512 and 640 shader versions...)

Otherwise, maybe get an old Bonaire card? You can find the various models and specs, here:

https://en.wikipedia.org/wiki/List_o..._RX_400_Series

Do they work with ROCm? I'm confused because there is no good list of GPUs or GCN generations that can be used with ROCm. They say that Fiji, Polaris and Vega are supported, which makes it hard to buy anything but very latest hardware (which might be a goal though ).

Those are toys, IMO.

If you're prototyping an embedded solution and need a way to test/evaluate the inferencing performance of that chip, or if you need to add some inferencing horsepower to a Raspberry Pi class machine, then go for it. Otherwise, any real GPU should kick them to the curb.

Okay. I really don't have an idea what those devices are meant for really. I've only used deep learning on CPUs and GPUs so far.

[AndyChow]

When you say "current git" do you mean the repo that coder linked to (which builds and runs on ROCm AFAIK) or latest Tensorflow git ?

I wouldn't be surprised if latest Tensorflow git failed because of some new dependencies, but AFAIK we are running Tensorflow and other frameworks today.

The current Tensorflow git fails.

git+https://github.com/tensorflow/tensorflow

nccl problem, which is nVidia tech. I have AMD tech.