Call me back when ROCm supports their flagship graphics card.

The 5000 series has been out for an entire year without ROCm support at this point.

That is because AMD split graphics vs. compute and you look at a non-compute card. The Vega-based chips work quite well and the Radon VII is (or was, I'm afraid) an excellent value proposition.

So this is well outside my area of expertise; but that seems weird. I want AMD to shepherd the ecosystem to a point where I can run PyTorch with some support from a graphics card. Supporting some graphics cards and not others doesn't sound very promising.

It is amazing to watch how much of a struggle AMD is having with getting PyTorch to work with ROCm. It makes me appreciate what a good job Nvidia must have done with CUDA.

Nvidia has split lines too. But they understand the importance of getting developers on their platform.

(To be fair AMD has it planned, but it's behind several other priorities to them.)

> Vega-based chips work quite well

LOL what a fucking joke. Not even close.

Also that one VFIO bug makes one of the great advantages of Radeon disappear and there is no fix in sight.

Fuck AMD. Bunch of marketing hype and they ship garbage half baked crap to the market.

Yeah... I learned this the hard way.

I thought AMD had great Linux driver support, and open source drivers.

Turns out AMD only has the open source driver part of the things, and this driver does not support a large chunk of their products.

I'd take a binary driver over no driver any day (and in fact, I returned the amd card and am using intel now, which works great).

difference between AMD and nvidia:

AMD: some open source drivers, but many things are inop

Nvidia: all closed source, but everything works

Source: I have a Vega 56 and I am gonna give it the sledgehammer when 3000 series arrives. Fuck that shit to high heavens my dude. ROCm is buggy as fuck. Only the bare basics work on Linux. Drivers are buggy, and crash all the time. Even the most basic bullshit is not implemented right. Like there is no fan control. That's how bad it is.

ROCm exists but many frameworks don't support it.

https://ai.stackexchange.com/a/16541

PyTorch supports ROCm quite well if you can follow three step build instructions instructions (https://lernapparat.de/pytorch-rocm/). TVM.ai also supports ROCm well. TF I didn't check.

FP32 performance for the VII is comparable to GTX20280Ti e.g. on Resnet training. Tensor Cores are cool, but 90% of people don't use FP16 that much.

I'm not sure that is the best reply that stackoverflow has ever seen.

The problem is that ROCm does not support many AMD compute cards, like their top of the line Navi-based one-year-old 5700 family.

Most frameworks do support ROCm though, so the ball is on AMDs court.

I can only speak for myself, but my understanding is that the Navi cards are not intended nor advertised for compute. Now, one might wish to get a consumer card with Arcturus after the Instinct MI100 is released or dislike that the Radeon VII is discontinued, or even that it wasn't split into render and compute lines, but basically, to me the complaint about Navi not being supported by ROCm sounds a bit like buying a boat and complaining that it can't fly. Driver support was bumpy for me with Linux 5.0 or so, but I haven't had trouble with just running whatever kernel Debian ships for quite a while now. But I must admit I don't use the GPU for anything but compute, so I don't know about graphics.

But your mileage may vary, I don't want to make people swear.

a GPU is a GPU, it's all just programs running on the shaders.

If AMD is unwilling to properly support its consumer GPUs in GPGPU workloads and NVIDIA is, then well... sounds like a good reason not to buy AMD.

It's really not a major ask, people have been running GPGPU programs on consumer cards since forever, even though they're "consumer" and sold as graphics cards and not compute accelerators. NVIDIA basically does this on purpose, people use their personal hardware to get a foot in the door of the ecosystem and end up writing programs that get run on big compute GPUs which NVIDIA makes big profits on, it's very intentional.

If AMD chooses not to do that, well... can't really blame people for avoiding their stuff when NVIDIA is willing to let you do this stuff on their cards and AMD isn't.