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

[phoronix]

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

Hitting mainline LLVM and Clang compilers today were support for Vega 12 "GFX904" and Vega 20 "GFX906" graphics processors...

http://www.phoronix.com/scan.php?page=news_item&px=Vega-12-20-AMDGPU-LLVM

[valici]

WTF is Vega12?. It doesn't appear in AMD's roadmap for this year... So unless it's a boring workstation card, it's gonna be a surprise.
Hoping it's the second one..

[paupav]

It would be really bad naming scheme if vega 12 wasn't workstation card.

[Raven3x7]

What's the point of Vega 20 if there's no Tensorflow support? Hopefully that will get sorted soon

[coder]

WTF is Vega12?. It doesn't appear in AMD's roadmap for this year... So unless it's a boring workstation card, it's gonna be a surprise.
Hoping it's the second one..

Old news. We have received clarification that the name indicates nothing more than the sequence in which the chip was designed. So, no clue as to how big it is.

We also know it's not Vega M, since that actually appears to be Polaris-based.

It would be really bad naming scheme if vega 12 wasn't workstation card.

No, why? AFAIK, AMD doesn't even use the word "Vega" in their workstation cards.

Okay, there's Vega Frontier, but that was a weird sort of pro-sumer hybrid. Their main Vega 64 workstation card is the AMD Radeon Pro WX 9100. To be honest, it feels like the biggest different is probably ECC support in the Pro.

Anyway, AMD doesn't normally make completely new chips just for workstation. They follow the industry standard practice of using the same GPUs in both consumer and workstation products. It's only the very largest GPUs that fail to reach mass market, such as Nvidia's P100 and V100 chips. Vega 20 looks to follow this path.

[coder]

What's the point of Vega 20 if there's no Tensorflow support? Hopefully that will get sorted soon

*ahem*

https://github.com/ROCmSoftwarePlatform/tensorflow

[coder]

... sdot8, and udot8 ...

Okay, so we're now doing 4-bit ops? I guess they had to one-up Nvidia somehow.

...but the lack of any instructions for some kind of tensor-product is a little worrying. Maybe there's another way of feeding its tensor cores (assuming it has them). Perhaps they resigned to just building an inferencing-focused chip.

[smitty3268]

Hmm, I thought the Vega 12 cards were just the 540 and 550 cards, but it appears those are called Polaris 12, not Vega 12.

[coder]

Hmm, I thought the Vega 12 cards were just the 540 and 550 cards, but it appears those are called Polaris 12, not Vega 12.

Yeah, bridgman clarified that the number (in the code name - not the launched product name) simply refers to the order in which the chip was designed.

I think you could map the name to the major version of the architecture, the first digit to the minor version, and the last digit is a sequence number.

[AndyChow]

*ahem*

https://github.com/ROCmSoftwarePlatform/tensorflow

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.

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.