Huawei needs far more chips than Nvidia for the same computing power, meaning potentially more wafers from Semiconductor Manufacturing International Corp. and, crucially, far more memory. Matching Rubin’s compute would require about 1,180 terabytes of memory using Huawei chips, compared with just 138 terabytes for Rubin. Also, Huawei’s new innovation — near-packaged optics — moves the optical engine from the edge of the board to sit beside the chip — cutting power and delay. It uses fewer components but more optical fiber, making it a narrower win for Chinese optical suppliers.
메마겟돈