AMD Announces The Radeon Pro VII

[bridgman]

FWIW, nobody has ever said that RDNA is "bad at compute"... it's just that CDNA offers a compute-per-unit-area advantage over RDNA (and the competition IIRC), which is particularly important for datacenter compute.

[tomtomme]

Being a good architecture for compute means nothing if AMD says the opposite and doesn't provide the drivers. Has everyone forgotten Terascale already? Even after a decade, OGL4 support is still blocked because AMD says so. Being open source changes nothing, AMD controls that part of Mesa. GCN has been out for more than what Terascale ever got. It can be shut down tomorrow. Even today. Just to boost RDNA sales. They don't even need a valid reaso.

There are no professional compute products with RDNA and there won't be, because RDNA is bad at comput. AMD does _not_ say the opposite. Shutting down GCN would mean to abandon the compute market.

[eydee]

Being a good architecture for compute means nothing if AMD says the opposite and doesn't provide the drivers. Has everyone forgotten Terascale already? Even after a decade, OGL4 support is still blocked because AMD says so. Being open source changes nothing, AMD controls that part of Mesa. GCN has been out for more than what Terascale ever got. It can be shut down tomorrow. Even today. Just to boost RDNA sales. They don't even need a valid reason.

[nranger]

Making a pro version of an already failed product which is also on the verge of being artificially moved to artificial legacy state due to being a too old architecture.

What could go wrong? Poor Volta I guess.

Radeon VII was a poor gaming product, but for compute it was awesome. TFLOPS per die area was very impressive. As the last generation of desktop GCN, the architecture had bottlenecks in geometry processing and instruction scheduling when shader instructions couldn't fill a wave in the pipeline. But those issues go away when you're not trying to draw triangles or pixels on a screen.

There's a reason GCN is evolving in to "CDNA". Keep the big horsepower for double, half, and quarter precision in the HPC/Workstation products, go even bigger on compute/die area density. And hopefully that lineage means that AMD can continue to improve the software, since they are SO far behind CUDA as an ecosystem.

[tomtomme]

Making a pro version of an already failed product which is also on the verge of being artificially moved to artificial legacy state due to being a too old architecture.

What could go wrong? Poor Volta I guess.

for gpgpu / workstation workloads this card is very capable. only the software stack makes it suitable only for a niche. the architecture albeit is very good for compute / pro customers. it only fails for gaming.

[eydee]

Making a pro version of an already failed product which is also on the verge of being artificially moved to artificial legacy state due to being a too old architecture.

What could go wrong? Poor Volta I guess.

[tomtomme]

Maybe first GPU with Chiplets?

why would you think that? the chip is already a well known one and not that big, that chiplets even would make sense...

[CochainComplex]

A few differences I noticed:
1. TDP: 295W->250W, likely due to the improvements from TSMC
2. FP64: 3.5TFLOPS -> 6.5 TFLOPS, although I'm not sure FP64 is still a thing for GPU computing...
3. Infinity fabric for GPU P2P.

Maybe first GPU with Chiplets?

[seesturm]

No, PRO cards don't support SR-IOV.

[drlamb]

You could has 6 a while ago. Some of the FirePro W9000 series has 6.


Speaking of which, is that the only difference between this and the Radeon VII? Also... wasn't the Radeon VII basically just a binned workstation GPU? If so, does that mean this is a bin of a bin?

No. It supports PCIe 4.0 (3.0 on the VII) and has 1/2 rate Double Precision vs 1/4th on the VII. I wonder if it supports SRI-OV...