3 ms·
GDDR6, 14 GT/s if we want to be accurate. So about double the DRAM bandwidth, but less than 1/20th the SRAM capacity, and fairly different architectures on-chip
by wtallis 3y ago
GDDR6, 14 GT/s if we want to be accurate. So about double the DRAM bandwidth, but less than 1/20th the SRAM capacity, and fairly different architectures on-chip. I'm not sure what ML workloads would benefit from the extra SRAM enough to overcome the DRAM bandwidth deficit, but they probably exist or Tenstorrent wouldn't be making such a SRAM-heavy chip.