MP 7,1 If the 6900 XT is almost as fast as the Vega ii DUO, what will MPX releases look like?

[deconstruct60]

Just saw this about AMD CDNA2 with 2 Gpus in one, much like the duo. I wonder if this will make it to the Mac Pro too

AMD confirms Aldebaran CDNA2 GPU for Instinct MI200 has 'primary' and 'secondary' die - VideoCardz.com

AMD Aldebaran CDNA2 GPU features two dies In the most recent Linux kernel update, AMD engineers have confirmed that the upcoming CDNA2 GPU codenamed Aldebaran will feature two dies. While consumer graphics processing units are not expected to feature MCM (Multi-Chip-Module) design till at least...

videocardz.com

Probably not because Apple has so heavily weaves "shader language" Metal into GPU drawing Metal. ( Metal puts a shader skewed view on the GPGPU compute. It isn't driving to more general purpose. )

CDNA gains die space by dumping all of the display output logic ( no display driver units, some raster aspects, etc. ) .

If it was a "card' it would be more like an Afterburner card that was a computational accelerator only. That might help with the "all non Apple drivers out of the kernel" problem.

If Apple and AMD put in the driver effort they could do it. The question is do they want to ( is Apple willing to pay to get it done). However, there are still some apps that still are structured to presume the compute cores are on the same VRAM instance as the display draw driving ones.

Also if Apple is putting money into RX 6900 - 6800 like updates ( that lack Inifinity Fabric) then that also somewhat mutes the utility of CDNA. Especially if capped on a PCI-e v3 backplane. ( getting results to/from the CDNA card)

The primary solfware stack that CNDA is coupled to is somewhere where Apple isn't going. Stuff with an "Open" prefix ( OpenCL ) or cross platform ... Apple isn't interested in. Double floats ... fingers in ears , doesn't exist at Apple (not good for an smartphone GPU) . etc. etc.

 

[deconstruct60]

P.S.

Also if Apple is putting money into RX 6900 - 6800 like updates ( that lack Inifinity Fabric) then that also somewhat mutes the utility of CDNA. Especially if capped on a PCI-e v3 backplane. ( getting results to/from the CDNA card)

Not sure where AMD is going with the CDNA2 though. If AMD had multiple entries in that line up then perhaps AMD could create a "less bigger" die pair. Cut down from the max core count and drop the display controller/raster aspects either back into the GPU core die and/or interface I/O module. For example not two exactly sysmetric dies, but one that does some core rollback for raster and one that that is still all "compute core" skewed. Lower the TPD back into the "high" from "very very high" range and place the package on a standard size , length GPU card.

The "card" listed in the videocardz article isn't an "open" server card, not an old-school card. that makes it harder to get the GPU in front of a wider set of developers to write code to.

It is a bit strange that AMD doesn't have a workstation card with high bandwidth backhaul channel connector.
Nvidia has them. For now AMD just needs to get some traction. [ and not like they have a huge pile of "extra" wafers to use up on that product variation either. ] So perhaps they have kicked that off into RDNA3 and/or CDNA3 . That is another way to "add in" a raster/display back to a CNDA card... coupling it with Infinity Fabric to a RDNA card and just copying data over to be displayed over the backhaul. ( it does work for ultra frame rate gaming but for compute-and-display workloads it can work reasonably ok. ). And Apple did lay down some substantve ground work to make that happen in Metal.

A CDNA variant with some display ability could draw Apple back in. That would uncork the major disconnect with the baseline assumptions that Metal makes.