Infrastructure atlas
NVIDIA/Blackwell

B200 Tensor Core

Blackwell’s data-center GPU for FP4/FP6/FP8 workloads, with a large HBM3e memory pool and fifth-generation NVLink.

GPUBlackwelltraininginferenceFP4
B200 Tensor Core hardwareHGX B200 baseboard reference

01 / Memory

180 GB HBM3e

Per GPU

02 / Bandwidth

Up to 8 TB/s

Per GPU

03 / Compute

FP4 / FP6 / FP8

Tensor Core modes

Precision & peak compute

PrecisionPublished peakBasis
FP4 Tensor18 PFLOPSPer GPU · with sparsity
FP8 Tensor9 PFLOPSPer GPU · with sparsity
FP16 / BF16 Tensor4.5 PFLOPSPer GPU · with sparsity

Peak FLOPS and TOPS are theoretical hardware figures. Sparse and dense modes, chip and system totals, and form factors are not interchangeable.

NVIDIA HGX B200 specification summary

Technical specification

Silicon & memory
Architecture
Blackwell
Memory
180 GB HBM3e
AI formats
FP4, FP6, FP8, FP16, BF16, INT8
HGX board
8 B200 GPUs · 1.44 TB HBM3e
Software
CUDA, TensorRT, NCCL, NIM
System & interconnect
Scale-up
NVLink 5 and NVSwitch
Rack systems
HGX B200 and GB200 NVL systems
Form factor
SXM / OAM-class server module
Best fit
Large-model training and inference