NVIDIA GB300 Grace Blackwell Ultra Desktop Superchip 252GB

This is the NVIDIA GB300 Grace Blackwell Ultra Superchip as configured in the DGX Station, NVIDIA's desktop machine, rather than the datacenter part of the same name. Compute is identical to a full Blackwell Ultra datacenter GPU: 15 PFLOPS of dense NVFP4, 5 PFLOPS of dense FP8, and every figure divides cleanly out of the GB300 NVL72 rack totals. What NVIDIA cuts for the desktop is memory. The Station carries 252 GB of HBM3e at 7.1 TB/s where a rack GPU carries 288 GB at 8 TB/s, alongside 496 GB of LPDDR5X for 748 GB of coherent memory in total. The two pools are not interchangeable in practice, because they differ by roughly eighteen times in bandwidth, so a model that fits in HBM behaves very differently from one that spills. Note the INT8 figure, which sits far below FP8 rather than matching it. That is deliberate on Blackwell Ultra, whose INT8 datapath was reallocated to NVFP4, and it is not a typo to be corrected upward. This row covers a single Blackwell Ultra GPU. The NVIDIA GB200 Grace Blackwell Superchip row covers two B200s, so its figures are pair totals and sit at roughly double the ones here for reasons of packaging rather than silicon.