BREAKING
Llama 3.1 405B Trained in 7.07 Min
0
min
Time-to-Train
0
GPUs
0
Racks
Fastest in MLPerf Training v6.0
0
%
Weak scaling
0
%
Cross-rack overhead
0
GB/s
NVLink per GPU
Inside vs Between Racks
Intra-rack
NVLink
●
5th-gen NVLink
●
1,800 GB/s per GPU
Inter-rack
MRC fabric
●
ConnectX-8
●
Spectrum-X Ethernet
Blackwell Scales in Production
AI NEWS BLITZ
Azure and NVIDIA just set a record training Llama 3.1 405B in seven minutes.