The third level cache is integrated into the interchiplet bus
Both AMD and Nvidia are reported to be actively working on multi-chip GPUs. Nvidia may have such solutions of the Hoper generation, and AMD may switch to chiplets as early as next year.
These may be video cards in the near future. AMD Patent Describes Chiplet GPU With Unusual Cache
And today, for the first time, we can take a look at the technical details of AMD’s development, obtained directly from the patent. As always, the patent does not guarantee the implementation of the technology described in it, but in this case, everything looks like what AMD is already using.
So, the main feature of this patent is indicated in its name – Active Bridge Chiplet With Integrated Cache. That is an active chipset bus with integrated cache memory. We can say that this is an evolution of the Infinity Cache technology, which debuted in the Radeon RX 6000 graphics cards.
The essence of AMD’s idea is to turn an interface that connects individual GPU chiplets, including an L3 cache.
AMD explains that any communication between the chiplets will be through the active bus, which will be required to access memory channels on the individual GPU chiplets. In doing so, instead of relying on separate chiplet caches, the entire active bus cache will act as a monolithic GPU cache. That is, the memory will be addressed as a single register.
Unfortunately, it is not yet clear when such GPUs might hit the market.
