inference/research
verifiedhigh · 0.85updated 2026-10-04

Inside NVIDIA Blackwell Ultra: The Chip Powering the AI Factory Era

blog

Fields

id
sources/bw-nv-blackwell-ultra-blog
Publisher
NVIDIA
Kind
blog
Published
2025-08-22
Accessed
2026-10-04
notes
NVIDIA blog on Blackwell Ultra (B300/GB300). Key facts: 160 SMs, 640 5th-gen Tensor Cores, 15 PFLOPS dense NVFP4, 256 KB TMEM per SM, 288 GB HBM3e at 8 TB/s, NVLink 5 at 1.8 TB/s bidirectional per GPU, NVLink-C2C at 900 GB/s. Blackwell Ultra adds 2x attention-layer acceleration (accelerated softmax) and 1.5x more AI compute FLOPS vs Blackwell. SM has 128 CUDA cores, 4 Tensor Cores, 256 KB TMEM, SFUs. Fifth-gen Tensor Cores optimized for FP8, FP6, NVFP4. Dual-thread-block MMA where paired SMs cooperate on a single MMA.

Notes

NVIDIA blog on Blackwell Ultra (B300/GB300). Key facts: 160 SMs, 640 5th-gen Tensor Cores, 15 PFLOPS dense NVFP4, 256 KB TMEM per SM, 288 GB HBM3e at 8 TB/s, NVLink 5 at 1.8 TB/s bidirectional per GPU, NVLink-C2C at 900 GB/s. Blackwell Ultra adds 2x attention-layer acceleration (accelerated softmax) and 1.5x more AI compute FLOPS vs Blackwell. SM has 128 CUDA cores, 4 Tensor Cores, 256 KB TMEM, SFUs. Fifth-gen Tensor Cores optimized for FP8, FP6, NVFP4. Dual-thread-block MMA where paired SMs cooperate on a single MMA.

Cross-references

This record references no other records.