Skip to content

NVIDIA · Blackwell · 2024

NVIDIA GB200. 186 GB HBM3e, 2500 TFLOPS FP16 tensor.

Grace and Blackwell on one superchip with unified memory, for frontier-model training.

Availability

Not in stock right now.

On demand instead
B200 $10.01/hr

The card

186 GB of HBM3e. At 8.0 TB/s.

Memory
186 GB
HBM3e
Bandwidth
8.0 TB/s
Memory to compute
FP16 tensor
2500
TFLOPS
Per node
8 GPUs
NVLink, 1000 W each

Workloads

What teams run on the GB200. Sized for this card.

405B PretrainingBF16 training state for 405B (about 6.5 TB) fits in one NVL72 rack
Frontier Multi-ModalVision + audio + text in one pass
Llama 405B InferenceFP8 weights (405 GB) fit across 4 GPUs (744 GB)
Reasoning Models at ScaleDeepSeek-R1 in FP8 (671 GB) across 8 GPUs

Full specs

NVIDIA GB200 specifications.

Memory

Memory
186 GB
Type
HBM3e
Bandwidth
8.0 TB/s

Compute

FP32
90 TFLOPS
FP16 tensor
2500 TFLOPS
INT8
3600 TOPS

Power and form

TDP
1000 W
Form factor
Superchip
Architecture
Blackwell
Released
2024

Node

GPUs per node
8
Interconnect
NVLink
Maker
NVIDIA

Alternatives

Alternatives to the GB200. Same account, different price and performance.

GPUMemory · FP16 tensorStands outPrice gapFrom
NVIDIA GB200This GPU186 GB HBM3e · 2500 TFLOPSNot in stock right now
NVIDIA GB300288 GB HBM3e · 2500 TFLOPS+102 GB memoryIn stock · $14.30/GPU-hr3-month term, prepaid in full
NVIDIA B200180 GB HBM3e · 2250 TFLOPS$10.01/hr
NVIDIA B300288 GB HBM3e · 2250 TFLOPS+102 GB memory$17.02/hr
AMD MI300X192 GB HBM3 · 1307.4 TFLOPS-48% FP16 tensor$4.85/hr

Questions

Renting the GB200. What buyers ask first.

Other GPUs

Browse every GPU by architecture.