17.5GWofcomputingwith100MWofpower
With only a 28 nm process, we enable:
- 330GB @ 2.1 PB/s DRAM on die that obsoletes HBM and interposers
- 14,438 tokens/s80B-parameter decode at FP8 on a single die
- 174× tokens/W25.8 mJ per token vs ~4.5 J for HBM4 GPUs
- Space-gradeIntrinsically rad-hard — cosmic rays leave no lattice damage
64 tiers stacking into one monolithic dieLogic + memory fused — drag to rotate
