AI silicon isnโt compute-bound anymore.ย Itโs memory-bound.
In several recent conversations (including around @ChipStartUK / Silicon Catalyst UK), one point kept coming up:
๐๐ ๐๐๐๐๐ฒ๐บ๐ ๐ฎ๐ฟ๐ฒ ๐ป๐ผ๐ ๐ฑ๐ผ๐บ๐ถ๐ป๐ฎ๐๐ฒ๐ฑ ๐ฏ๐ ๐บ๐ฒ๐บ๐ผ๐ฟ๐; ๐ผ๐ณ๐๐ฒ๐ป ๐บ๐ผ๐ฟ๐ฒ ๐๐ต๐ฎ๐ป ๐ต๐ฎ๐น๐ณ ๐๐ต๐ฒ ๐๐๐๐๐ฒ๐บ.
Memory scaling has effectively stalled at ~16nm. That changes everything, making the real problem now is:
- moving data efficiently
- managing bandwidth
- controlling energy per bit
You see the same theme coming through in SemiEngineering and EE Times, the energy cost of data movement is overtaking compute so we are optimising data low flow and not just logic.
Physical implementation execution becomes a real differentiator efficiently moving and managing data at silicon level.
Read more in this EETimes article – โ๐ช๐ต๐ ๐ ๐ฒ๐บ๐ผ๐ฟ๐, ๐ก๐ผ๐ ๐๐ผ๐บ๐ฝ๐๐๐ฒ ๐ช๐ถ๐น๐น ๐๐ฒ๐ฐ๐ถ๐ฑ๐ฒ ๐๐ต๐ฒ ๐๐๐๐๐ฟ๐ฒ ๐ผ๐ณ ๐๐โ