AMD’s Patented GPU Chiplet Design Is Practical And Simple, But It Is Completely Outstanding
AMD’s design for its recently patented GPU chiplet is beneficial because it is quite conventional, unlike competitors like Intel and NVIDIA’s chiplet designs. If AMD’s design is as good as it is in theory, then it will stop the need for updating software as we have to do now.
AMD is not new to chiplet technology. In fact, it has used chiplets design for CPU and that has shown amazing results.
Chiplets are smaller and less complicated chips that together perform amazingly well with powerful processors. Because of their high-ended performance, these designs are being worked on by chipmakers like NVIDIA and Intel too.
On 31st December 2020, when the world was getting ready to enter the new year, AMD published its first patent for this design, and it is not only simple and conventional but if it materializes as it is, then it will be quite revolutionary and future-altering!
AMD’s design for chiplets graphics card is probably going to address the issues of latency, effective programming, and easy implementation of parallelism. High latency between the chiplets and harder implementation of parallelism are some of the key issues that occur and that have made it difficult for the chipmakers to attempt developing this design module. However, AMD’s design is going to be a little different.
AMD’s multi-chip design mimics the simple, monolithic chiplet design.
To understand this hypothetical design, we have to understand that its main advantage is a passive crosslink connection between the GPU chiplets. Now, these two chiplets that have this passive crosslink connection between each other use dedicated conductors to access the traces of the crosslink. This helps them intercommunicate effectively. These conductor structures are attached to the circuit board. This helps them draw power and other signaling mechanisms.
The crosslinking connection is between two caches, L2 and L3, but everything beneath them, including the cores and caches L1 and L2, are separate but still connected as they have the perception of their separation from the L3 chiplet.
L3 cache and the GDDR memory that are above this crosslink are also shared between the chiplets.
According to AMD, the speed of compute units accessing local low-level cache is really fast, and this is the same speed that they will be able to achieve to access low-level cache on other chiplets too. Experts believe that if this claim turns out to be true, then software updating issues can be resolved and systems will no longer need updating.
NVIDIA and Intel are aiming for some complex design architecture for these GPU chiplets.
Intel is working on a Hopper architecture for its GPU chiplets design, incorporating the technology of Embedded Multi-die Interconnect Bridge (EMIB) and another interposer technology. This design will drive power through a new memory fabric and will have a system-accessible cache.
On the other hand, NVIDIA may be working on the ‘Non-uniform Memory Access’ technology for a four-chiplet design.