NVIDIA B200 SXM 180GB

The NVIDIA B200 is the mainstream Blackwell datacenter GPU and the part most neoclouds mean when they advertise Blackwell capacity. It carries 180 GB of HBM3e at 7.7 TB/s in an SXM6 socket, draws up to 1,000 W, and is built from two reticle-limit dies joined by a 10 TB/s die-to-die link that the software stack sees as one GPU. Fifth-generation Tensor Cores add NVFP4, a microscaled 4-bit format, which is where the generational headline comes from: 9 PFLOPS of dense FP4 against the H100 SXM's 1,979 TFLOPS of dense FP8, with a second-generation Transformer Engine choosing precision per layer. NVLink 5 gives 1.8 TB/s per GPU, double Hopper's, which is what makes 72-GPU coherent domains practical. The number to be careful with is memory. NVIDIA announced the B200 at GTC 2024 with 192 GB and a higher FLOPS ladder; the shipping product is 180 GB, and a great many third-party spec tables still carry the announcement figures. The 180 GB recorded here is the shipping part, and NVIDIA's own DGX B200 datasheet corroborates it at 1,440 GB across eight GPUs.