NVIDIA B300 Blackwell Ultra 
Specification NVIDIA B300
Product Name NVIDIA B300 Tensor Core GPU
GPU Architecture NVIDIA Blackwell Ultra
GPU NVIDIA B300
GPU Memory 288GB HBM3e
Memory Interface 8192-bit
Memory Bandwidth Up to 8 TB/s
CUDA Cores 23,552
Streaming Multiprocessors (SMs) 160
Tensor Cores 640, 5th Generation
RT Cores 160, 4th Generation
Dense FP4 Tensor Performance Up to 15 PFLOPS
Dense FP8 Tensor Performance Up to 7.5 PFLOPS
Dense FP16/BF16 Tensor Performance Up to 3.7 PFLOPS
FP32 Performance Up to 75 TFLOPS
FP64 Performance Up to 37.5 TFLOPS
NVLink 5th Generation NVLink
NVLink Bandwidth 1.8 TB/s
Form Factor SXM
TDP Up to 1,400W
Memory Type HBM3e
ECC Yes
MIG Yes
Cooling Liquid / Server Cooling
Display Outputs None
Typical Platform HGX B300 / DGX B300
GPUs per System 8 × B300
Target Market Data Center / AI / HPC
Main Applications LLM Training, AI Inference, Generative AI, HPC