Actually, you could be wrong about that.... While prowling the AMD patents, something I do regularly, I came across a recent one detailing something called "GPU masking". Now, in the patent, I believe it had outlined the use of this technique on multi-gpu MCMs, which is basically taking the strategy used on Ryzen and applying it to GPUs.
The secret to it is that the system and API (DirectX) sees the MCM GPU as a single, logical entity and if I remember correctly, it accomplished the same feat with multiple memory pools. I take this to mean that AMD has figured out a way to use a chiplet approach with GPUs without game programmers and the API itself having to program specifically for multiple GPUs. If this is the case, then AMD may be able to bust the GPU market wide open.
That said, I would imagine that the same technique could possibly be applied to multiple gpu/videocard setups somehow.
That is very cool. I too think amd is up to something & ur the first who seems on a similar wave length. - not because i am clever, but for the humble reason that i think lisa knows what she is doing better than most bloggers.
The multi gpu patent you describe, is of course a perfect fit for Infinity Fabric - as u ~say - its like transferring zen architecture to gpu.
Cache coherency is Fabric in a nutshell. AMDs focus on it is at the root of their success, so the patents you describe dont surprise me. Thhat is exactly where i think they would be trying to go.
clearly we have a task (graphics) too big for a single gpu, but we are at an enduring impasse in teaming multiple processors (sli & crossfire ~fails)... multiple cheap, easily cooled & efficient gpuS would have more drastic effect on gpu than zen did on cpu.
Specifically , (I am not competent in gpu tech, but) i have long suspected that Vega was prefered for renoir for secret reasons - not the timing factors officially stated.
There seem to be apps (scientific/math e.g.?) where Vega is preferred. Maybe its more suited to some even more tempting prize than better consumer gaming?
AI is changing the usual processing paradigms - the AI raw data is potentially so vast that consequent slow costly transmission any distance, make processing by mini nodes on the edge of the data storage much more attractive - ~decentralising.
Maybe banks of tightly integrated hybrid processor APU's are suited, & can form a big new market to add to their already broad appeal?
The patents would fit amd's MO very well - they love serving multiple markets with easily scaled variants of a few standard ingredients - or even better win a new tier or market w/ ~existing recipes. (the 3900 & 3950 12 & 16c zens paired w/ x570 mobos invaded, a big patch of workstation turf using desktop cpuS recently). Similarly, 64c & 2x ram TR has charged upscale into ~epyc turf.
) so you shouldn't expect the same push for advancement, it's all perfectly logical. Like that guy, you don't, or don't want to understand that the segment in which these things exist has many constraints that prohibit significant leaps in performance of the same order compared to other segments.