NVIDIAEstimated

B300 (Blackwell Ultra)

Blackwell · Blackwell Ultra · SXM · 2025

Blackwell Ultra: 288 GB HBM3e with a further FP4/FP8 uplift over B200, aimed at the largest training and long-context inference. Specs still firming up.

LLM trainingLLM inferenceMultimodalHPC
Precision fingerprint
6432t3216BF168i8
Memory

288 GB

HBM3e

Bandwidth

8 TB/s

peak

TDP

1400 W

liquid

Max model

~120B

FP16, planning est.

Compute throughput

FP6440 TFLOPS
FP3240 TFLOPS
TF321.20 PFLOPS
FP162.50 PFLOPS
BF162.50 PFLOPS
FP85.00 PFLOPS
INT85.00 PFLOPS

Platform & software

InterconnectNVLink 5 — 1.8 TB/s
PCIePCIe 6.0 x16
Coolingliquid
MIGSupported
PartitioningMIG
VirtualizationMIG, vGPU
FrameworksCUDA, TensorRT-LLM, Triton, NeMo, vLLM
Availabilitybare-metal
Known limitations
  • ·Blackwell Ultra — public specs still emerging; figures are estimates
  • ·≈1.4 kW demands direct-liquid cooling and rack redesign
Related GPUs
NVIDIA

H100 SXM5

80 GB · 989 TFLOPS FP16

NVIDIA

H100 PCIe

80 GB · 756 TFLOPS FP16

NVIDIA

H100 NVL

94 GB · 989 TFLOPS FP16