AMD’s multi-chiplet GPU design would possibly lastly come true

An attention-grabbing AMD patent has simply surfaced, and though it was filed some time again, discovering it now could be all of the extra thrilling as a result of this tech is likely to be nearer to showing in future graphics cards. The patent describes a multi-chiplet GPU with three separate dies, which is one thing that might each enhance efficiency and in the reduction of on manufacturing prices.

Within the patent, AMD refers to a GPU that’s partitioned into a number of dies, which it refers to as GPU chiplets. These chiplets, or dies, can both perform collectively as a single GPU or work as a number of GPUs in what AMD refers to as “second mode.” The GPU has three modes in complete, the primary of which makes all of the chiplets work collectively as a single, unified GPU. This permits it to share sources and, as Tom’s {Hardware} says, permits the front-end die to take care of command scheduling for all of the shader engine dies. That is just like what a daily, non-chiplet GPU would do.

The second mode is the place it will get attention-grabbing. On this mode, each chiplet counts as an impartial GPU. Every GPU handles its personal job scheduling inside its shader engines and doesn’t intrude with the opposite chiplets. Lastly, the third mode is a mixture of the 2, the place some GPUs perform as their very own entity whereas others mix the chiplets to perform collectively.

As talked about, this patent is just not new. It was filed on December 8, 2022, simply after AMD launched the RX 7900 XTX and the RX 7900 XT. Though leakers have predicted that AMD would possibly go down the multi-chiplet route for a minimum of a technology or two now, this structure is at present solely actually utilized in AMD’s information heart GPUs. AMD has already dipped its toes in related tech in RDNA 3, although, with a design that used a graphics compute die (GCD) and a number of reminiscence cache dies (MCMs) for the reminiscence interface.

There are tangible advantages to switching to one of these structure, as per the patent: “By dividing the GPU into a number of GPU chiplets, the processing system flexibly and cost-effectively configures an quantity of lively GPU bodily sources based mostly on an working mode.” If it may turn into cheaper to supply these kinds of GPUs somewhat than utilizing more and more bigger monolithic dies, we’d begin seeing this design exterior of the information heart and within the GPUs all of us use in our personal computer systems.

Early leaks about RDNA 4 graphics cards teased AMD going with a full multi-chiplet design, and it’s straightforward to think about that the ultimate consequence may have resembled what we see within the patent. Nevertheless, with the information that AMD is sticking to midrange graphics playing cards on this subsequent technology, any hope of a multi-chiplet GPU appears misplaced for now. Maybe we’ll see this design come to life in RDNA 5.