Ahead-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 knowledge middle GPUs. The corporate’s newest RDNA 3 PC graphics playing cards even incorporate fundamental chiplets. Nevertheless, 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. Every 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 varied configurations.
As an example, 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 unbiased 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. As well as, 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 will be assembled utilizing a small variety of tape-outs and a multiple-die GPU will be constructed out of GPU chiplets that implement various generations of know-how.”
This chiplet design might additionally provide value optimizations through the use of completely different dies produced on a mixture of course of nodes. Solely essentially the most essential elements, like shader cores, should be fabbed on costly modern course of applied sciences. The supporting frontend logic might stay on cheaper, older silicon.
After all, AMD has already adopted rudimentary variations of this chiplet philosophy for his or her present RDNA 3 GPUs. Nevertheless, these are comparatively easy designs utilizing two kinds of chiplets – a main 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 sequence.
Nevertheless, the patent seems to take issues additional by enabling varied chiplet mixtures and configurations. It is unclear precisely when or the place we’d see this patent flip into reality–not all patents see the sunshine of day. Nonetheless, it aligns with current business tendencies that time to a transition to disaggregated chip designs. Pink Workforce has experimented with different distinctive chiplet designs. A more moderen patent filed in August 2023 describes a multi-die GPU design with no central processor directing the assorted chiplets.
