NVIDIA B300 SXM 262GB

The NVIDIA B300 is the Blackwell Ultra refresh of the B200, and it is a memory and inference part rather than a general uplift. Capacity rises to 262.5 GB of HBM3e at 8 TB/s on the air-cooled HGX and DGX boards this row describes, or 288 GB in the liquid-cooled GB300 NVL72 rack, board power rises to 1,400 W, and dense NVFP4 rises by half again to 13.5 PFLOPS while sparse NVFP4 stays where the B200 left it. That asymmetry is the whole design: Blackwell Ultra spends its transistor budget on dense low-precision throughput and on attention, which NVIDIA rates at twice the B200's, because reasoning models generate long chains of tokens and are bound by decode rather than by training arithmetic. The trade is visible in what got cut. INT8 is not a serious datapath on this part, sitting far below FP8 because the silicon behind it was reallocated to NVFP4, and FP64 is similarly de-emphasised, so a B300 is a poor choice for classical HPC in a way a B200 is not. Read those two rows as deliberate, not as errors.