Existing Confirmed

NVIDIA Coreweave MLPerf v3.0 Submission 2023

NVIDIA,CoreWeave,Inflection AI • United States of America • June 1, 2023
Total GPUs
3.6K
H100 Equiv
3.6K
Power
5.2 MW
FP16
3.5K PF

System Equivalents

This datacenter's 3.6K GPUs equal:

NVIDIA DGX H100
8 GPUs/system
448
NVIDIA HGX H100 (8-GPU)
8 GPUs/system
448
NVIDIA HGX H100 (4-GPU)
4 GPUs/system
896

Specifications

Primary GPUNVIDIA H100 SXM5 80GB
GPU Count3.6K
SectorPrivate

Additional Information

The 10,752 H100 GPUs far surpassed the scaling in AI training in June, when NVIDIA used 3,584 Hopper GPUs.

Inflection lent 3,500 H100 GPUs to Corweave (who appears to be the party that built that cluster) for the June MLPerf submission. Given that this is just slightly larger and operational just after, it appears very likely that this cluster is an extension of the MLPerf cluster (or the MLPerf cluster was just a subset of this cluster). It's possible that both of these subsets of the Eos-DFW cluster, but there isn't evidence for this besides that Eos-DFW was submitted to MLPerf later that year