AMD Announces The Radeon PRO W7800/W7900 Series

[phoronix]

Phoronix: AMD Announces The Radeon PRO W7800/W7900 Series

As the "world's first pro chiplet GPU", AMD today is announcing the Radeon PRO W7000 series as their first RDNA3-based professional offerings.

https://www.phoronix.com/review/radeon-pro-w7000-series

[darkbasic]

Does it support hardware virtualization or it's still not expensive enough?

[dimko]

Does it support hardware virtualization or it's still not expensive enough?

isn't PCIE GPU virtualization a functionality of a motherboard?

[Radtraveller]

$3500+ ??

[ms178]

Which parts of the professional market are they targeting with these? I thought CUDA still rules wide portions of the professional market. Also I wonder why they are targeting that market with RDNA3 and not the latest CDNA architecture.

[nyanmisaka]

No proper ROCm support making this and RX 7900 a piece of crap.

Users have been begging ROCm team to support the RX 7900 for half a year and nothing has happened.

https://github.com/RadeonOpenCompute/ROCm/issues/1880

[stormcrow]

isn't PCIE GPU virtualization a functionality of a motherboard?

Motherboard chipset, firmware, and to lesser degree host OS support the various interfaces needed for direct virtualization pass-through, so I'm given to believe.

Which parts of the professional market are they targeting with these? I thought CUDA still rules wide portions of the professional market. Also I wonder why they are targeting that market with RDNA3 and not the latest CDNA architecture.

Nvidia has most of the GPU compute support sewn up. That said, there's blood in the water with the huge furor over supporting the next generations of LLM compute models. So while Nvidia is the gorilla incumbent, it's not necessarily a given they will remain so if another market player offers better performance and tools tailored for those market segments. Also, there's another compute model in the works that's kinda buried in the noise right now. High performance homomorphic encryption requires hardware that's currently not available commercially but supposedly on the drawing board. Whoever gets to that gold vein first is going to be the next hardware tech darling.

Edit to add: What I'm getting at is with language models there's now a tools break where the traditional HPC tools aren't entirely suited for software that utilizes neural processors. So there's opportunity for Nvidia to drop the ball in the new market segment. Right now, even though PCs don't normally come with a neural processor on the client side, that's not going to be the case going forward. All Apple M processors have a neural processor. Future PCs will eventually have a neural processor of some kind on the CPU if not the GPU. That means developers need tools and they need hardware to develop those tools and models. That means the market is not currently set on a single vendor like traditional HPC is with Nvidia actually before CUDA itself existed. CUDA was a response to GPGPU that was being done at the assembly level with mostly Nvidia GPUs. Well, right now there's a similar opportunity in the neural processors and in a few years a shift in encryption computing models (which won't need quantum computing).

[darkbasic]

Motherboard chipset, firmware, and to lesser degree host OS support the various interfaces needed for direct virtualization pass-through, so I'm given to believe.

I'm not talking about passthrough but GPU sharing across multiple guests.

[ripper81]

No proper ROCm support making this and RX 7900 a piece of crap.

Users have been begging ROCm team to support the RX 7900 for half a year and nothing has happened.

https://github.com/RadeonOpenCompute/ROCm/issues/1880

Which is not unusual at all. ROCm is designed to run on selected professional GPU'S. Developement is focused on the datacenter. Support for consumer GPU'S (Radeon) is unofficial. However since AMD has now launched two professional RDNA3 GPU's there is hope for support in ROCm. This could enabled unofficial support for the RX 7000 GPU series as well.

[MorrisS.]

I'm waiting for entry level GPUs. I hope this series will provide any of them.