BREAKING
GB300 NVL72: Up to 25x Per Watt
Perf Per Watt vs Hopper
DeepSeek V4 Pro
25
GLM 5.1
20
Kimi K2.6
10
0
x
throughput per MW gain
0
GPUs over NVLink
Why Per Watt Matters Now
The Pitch
NVIDIA
●
Lower cost per token
●
Better operator margins
●
Benchmark to production
The Caveats
Observers
●
Limited independent testing
●
Numbers depend on workload
●
Heavy cooling demands
0
kW
peak rack draw
0
target PUE
Efficiency Becomes The Metric
AI NEWS BLITZ
NVIDIA says its GB300 NVL72 delivers up to twenty-five times better performance per watt than Hopper.