The Next generation of Nvidia GPUs are not even out yet and the rumors for even the next GPU architecture is already in the wild. The architecture is code named Hopper and it will totally change the current design of the GPU. At present, every GPU by every vendor has a single monolithic design. but Hopper is set to change this by using MCM design.
What is MCM design?
This is called MutiChip Module. Right now we have a single GPU in a single package, the GPU has to be huge to get more transistors which can be shrunk down with small manufacturing processes. The system identified this system as one GPU. With MCM design there can be multiple GPU in a single package that can work in tandem.
Think of this like AMD’s Chiplet design on CPU side. Where multiple chips are packaged in single die and its like one giant CPU.
Why MCM design?
There is a limit to how much a GPU can be made larger. Also in one large silicon, if there is a defect, the whole silicon goes to waste. but with small GPUs glued together, any silicon with the defect can be discarded which will be a small wastage. Further, the GPU silicon can be stacked in 3D instead of side by side which can also save space.
So why didn’t we have MCM design before?
Well, we already have MCM design. Its called the chiplets in CPU world. AMD pioneered it and now Intel has also announced its Chiplet CPUs. In GPU world we had SLI and Crossfire which although technically not MCM design, but did try to use multi GPUs (and failed miserably). What we didn’t had was MCM GPU. That is because:
- Every software (i.e. games and allied APIs) all work on single GPU. Thus it is very difficult for single MCM design to pretend its a single GPU. Thus every software and API have to be reimplemented to proper utilize multiple GPU
- Interconnect speed: In case of multiple GPU designs, the silicon has be to be connected with each other. This interconnect speed is kind of slow compared to one monolithic design where one die has all the transistors. This now can be solved with high speed interconnect solutions available.
AMD was contemplating MCM design for Navi but was discarded. David Wang (Sr vice president Engineering) had the following to say about MCM design
To some extent you’re talking about doing CrossFire on a single package. The challenge is that unless we make it invisible to the ISVs [independent software vendors] you’re going to see the same sort of reluctance. We’re going down that path on the CPU side, and I think on the GPU we’re always looking at new ideas. But the GPU has unique constraints with this type of NUMA [non-uniform memory access] architecture, and how you combine features… The multithreaded CPU is a bit easier to scale the workload. The NUMA is part of the OS support so it’s much easier to handle this multi-die thing relative to the graphics type of workload
At the time of the launch of Hopper GPU series, Nvidia will be able to utilize 5nm manufacturing node. Further, the interconnect system will improve further. Nvidia already has NVlink in their arsenal. With these, they are perfectly capable of launching MCM design. Who knows they might deploy ray tracing core as separate package altogether. With Intel entering the market, The GPU was is not going away anytime soon.
- Blender 3D Jumps to Version 3.1 with Massive Changes Baked In - March 10, 2022
- Open Source Painting and Illustration App Krita 5.0.0 Released- Faster with Massive Feature Updates - December 23, 2021
- Unity Completes Acquisition of Weta Digital- $1.625 Billion spent well - December 6, 2021