En vivo
AI hardware

NVIDIA GB300 (Blackwell Ultra)

FP16 throughput
2.5×10¹⁵ FLOP/s
Memory
268 GB
TDP
1.4 kW

NVIDIA GB300 (Blackwell Ultra) is an AI accelerator designed by NVIDIA (GPU), released in August 2025. It is built on a 4 nm process at TSMC.

It delivers 2.5×10¹⁵ FLOP/s of tensor FP16/BF16 throughput, 5×10¹⁵ FLOP/s at FP8 and 1.5×10¹⁶ FLOP/s at FP4. It carries 268 GB of memory at 7.3 TB/s of bandwidth. Its thermal design power is 1.4 kW.

Full record
Manufacturer
NVIDIA
Type
GPU
Release date
Aug 22, 2025
Tensor FP16/BF16
2.5×10¹⁵ FLOP/s
FP8
5×10¹⁵ FLOP/s
FP4
1.5×10¹⁶ FLOP/s
FP32
8×10¹³ FLOP/s
FP64
1.3×10¹² FLOP/s
INT8
1.6×10¹⁴ OP/s
Memory
268 GB
Memory bandwidth
7.3 TB
TDP
1.4 kW
Process node
4 nm
Foundry
TSMC
More from NVIDIA
SourceEpoch AI, 'Machine Learning Hardware'. Published online at epoch.ai. Retrieved 2026-07-29 from https://epoch.ai/data/machine-learning-hardware. Licensed under CC BY 4.0.
← All ai hardware