Nvidia evaluating multi chip module GPU designs

[HEXUS]

Green team reckons that MCM architecture can counteract Moore's Law slowdown.

Read more.

[Gunbuster]

So kind of like Ryzen and Infinity Fabric. Looks interesting.

[Goodman2576]

already happened once. Hope this time will end much better:
http://www.tomshardware.com/reviews/gigabyte,954-2.html

[Tabbykatze]

So kind of like Ryzen and Infinity Fabric. Looks interesting.

That is exactly what AMD are hoping to achieve with Infinity fabric, as well as using it for better CPU/GPU interconnects.

[HW90]

already happened once. Hope this time will end much better:
http://www.tomshardware.com/reviews/gigabyte,954-2.html

It's a bit different to that, it's more akin to how threadripper uses 4 CPUs for one big SoC than say the GTX295 which has two different SoCs.

[Gunbuster]

Yeah basically IF can be internal, ie between the two 4core blocks in Ryzen or external say between 2 on Threadripper, 4 on Epyc or between the sockets in a dual Epyc board.
Nvidia's GPUs are already pretty modular internally, by making the connection more flexible between the clusters would allow for easier to scale architecures. ie, a 2080 could just be two 2060s on one chip.

FYI my language is muddled as I am a layman in these kind of things.

[scaryjim]

It's a bit different to that, it's more akin to how threadripper uses 4 CPUs for one big SoC than say the GTX295 which has two different SoCs.

Threadripper uses 2 dies EPYC uses 4.

And if you read the paper this is actually a little different. Threadripper/EPYC are full CPU dies tied together on a fabric. The proposed MCM GPU is not full GPU dies. The slices contain shaders / caches / memory channels, but the whole thing requires a separate control block - a distinct and separate die - to handle thread dispatch etc. Also the MCM approach doesn't net significant benefits over an optimised multi-GPU setup until you've gone through some very stringent optiimsation processes, and those optimisations only work for about half of all GPU workloads, and actually hurt performance in some (we've already discussed this a week ago in http://forums.hexus.net/cpus/371038-...ml#post3827056).

Managing the extreme degrees of parallelism in a GPU makes it far more complicated to optimise multi-GPU or even multi-die shaders slices than it does with multiple CPUs. That's why we've had MCM CPUs for a long time, but not MCM GPUs....