Forward-looking: Chipmakers are getting critical about splitting up complicated chip designs throughout a number of smaller “chiplets.” AMD is main the cost, having already carried out multi-chiplet architectures for its CPUs and information middle GPUs. The firm’s newest RDNA 3 PC graphics playing cards even incorporate primary chiplets. However, a brand new patent submitting exhibits these designs haven’t realized their full potential.
The December 2022 patent describes a GPU design cut up throughout a number of GPU chiplet units. Each chiplet has a frontend die paired with a number of different shader-engine dies. The intelligent half is that these GPU chiplets can flexibly mix in numerous configurations.
For occasion, chiplet units can work collectively as a single, unified GPU, working like a standard monolithic GPU design–or AMD can cut up the units into distinct teams, every functioning as an impartial GPU. There’s even a hybrid mode the place some chiplet units act as a unified GPU whereas others function independently.
This modular GPU design has a number of advantages. For one, it permits scaling up or down GPU assets and efficiency based mostly on product wants or working modes.
“By dividing the GPU into a number of GPU chiplets, the processing system flexibly and cost-effectively configures an quantity of energetic GPU bodily assets based mostly on an working mode. In addition, a configurable variety of GPU chiplets are assembled right into a single GPU, such that a number of completely different GPUs having completely different numbers of GPU chiplets could be assembled utilizing a small variety of tape-outs and a multiple-die GPU could be constructed out of GPU chiplets that implement various generations of expertise.”
This chiplet design might additionally provide price optimizations through the use of completely different dies produced on a mixture of course of nodes. Only essentially the most essential parts, like shader cores, must be fabbed on costly modern course of applied sciences. The supporting frontend logic might dwell on cheaper, older silicon.
Of course, AMD has already adopted rudimentary variations of this chiplet philosophy for their present RDNA 3 GPUs. However, these are comparatively easy designs utilizing two forms of chiplets – a major GPU die, and reminiscence cache dies. Rumors point out that Nvidia is engaged on chiplets for the compute GPUs in its upcoming GeForce RTX 5000 collection.
However, the patent seems to be to take issues additional by enabling numerous chiplet mixtures and configurations. It’s unclear precisely when or the place we would see this patent flip into reality–not all patents see the sunshine of day. Still, it aligns with latest trade traits that time to a transition to disaggregated chip designs. Red Team has experimented with different distinctive chiplet designs. A more recent patent filed in August 2023 describes a multi-die GPU design with no central processor directing the varied chiplets.