ROCm doesn't completely suck with the Radeon VII, which is a Radeon Instinct MI50. Deep learning is not my day job and I'd like to avoid supporting NVidia's insane prices for DL-capable cards, so I've been dealing with the performance hit, and only using the R7 for DL tasks then switching it off when the power isn't needed. The XFX Radeon VII is actually on sale for Newegg for $569 so it's a lot of power (and 16 GB HBM2) for that price.

Agreed. The Radeon VII is currently the best price/compute GPU out there for deep learning. It's performance on RESNET-50 is about the same as the 2080Ti -

https://github.com/ROCmSoftwarePlatform/tensorflow-upstream/...

That's only theoretical. Try to use ROCm on latest frameworks or on external models that write custom CUDA operations/losses. Basic stuff might work in a more complicated way than on NVidia, advanced stuff is guaranteed to either not work or work in a couple of months when it lands into ROCm.

Radeon VII is a beast for FP64 computation, if you do simulation or heavy computations that require supercomputer-level of precision, then grab one while you can, it's the best price/performance of all GPUs on the market.

However folks, please don't follow the advice about using for it Deep Learning if you want to actually have Deep Learning business in any way.

Companies like https://myrtle.ai/ and https://www.graphcore.ai/ are popping up.

Nope, just grab yourself a RTX 8000 and be able to train latest SOTA models (albeit slowly). Titan RTX is already insufficient and nobody else is in the game for actually owning DL hardware :(

Google has TPUs now

But only through Google Cloud for now, as far as I'm aware.