According to the statements of the chief scientific adviser of NVIDIA Bill Dally (Bill Dally) in an interview with the resource , the company developed the technology for creating a multi-core processor with a multi-chip layout six years ago, but is still not ready to use it in mass production. On the other hand, the company also started placing HBM-type memory chips in close proximity to the GPU several years ago, so it cannot be blamed for completely ignoring the βchiplet fashionβ.
So far it has been argued that A 36-core RISC-V processor from NVIDIA was needed to test methods for scaling performance in computing accelerators, as well as to prepare for the introduction of new layout solutions. All this experience, according to representatives of NVIDIA, may be required by the company at a time when it becomes economically feasible to create graphics processors from separate "chiplets". Until such a moment has come, and NVIDIA does not even undertake to predict when this will happen.
Bill Dalli also noted that relying on lithography for scaling processor performance is no longer worth it. Between two adjacent stages of the technical process, the increase in the speed of transistors is measured by 20%, in the best case, and architectural and software innovations can increase the speed of GPUs by several times. In this sense, architecture dominates lithography from NVIDIA's point of view.
This position was repeatedly confirmed in his statements by the founder of NVIDIA Jensen Huang (Jen-Hsun Huang). Until now, he did his best to prove the progressiveness of the approach to creating monolithic crystals, spoke disparagingly about competitors who are chasing new technical processes, and even compared βchipletsβ as a joke with consonant chewing gum (βchicletsβ), explaining that he only likes latest interpretation of this term. Nevertheless, the statements of NVIDIA specialists closer to product development allow us to believe that the company will eventually switch to a multi-chip layout. Intel, for example, does not hide its intentions to make a 7nm GPU multi-chip using the Foveros layout. AMD is actively using "chiplets" when creating central processors, but in the graphics segment it has so far limited itself to "populating" HBM2 type memory.
Source: 3dnews.ru