China’s Dongfang Suanxin (DFSX) is on a quest to prove that memory bandwidth, and not an incessant obsession with node miniaturization, is the key to unlocking sizable economies of scale for AI workloads, as demonstrated by its DF1000 and the upcoming DF2000 chips, typically arrayed within TY64 SuperNodes to give NVIDIA’s GB200 NVL72 system a run for their money. A DFSX TY64 SuperNode, made up of 14nm DF2000 chips, will offer a memory bandwidth of 960TB/s vs. just 576TB/s for NVIDIA GB200 NVL72 systems For the benefit of those who might not be aware, DFSX’s DF1000 chip is fabricated using […]
Read full article at https://wccftech.com/chinas-dfsx-offers-2x-the-memory-bandwidth-of-nvidias-gb200-nvl72-system-with-a-14nm-supernode-that-skips-microbumps-for-vertical-compute-memory-towers/