Existing Likely

Oracle OCI Supercluster H100s

Oracle • United States of America • April 15, 2024
Total GPUs
16.4K
H100 Equiv
16.4K
Power
23.4 MW
FP16
16.2K PF

System Equivalents

This datacenter's 16.4K GPUs equal:

NVIDIA DGX H100
8 GPUs/system
2.0K
NVIDIA HGX H100 (8-GPU)
8 GPUs/system
2.0K
NVIDIA HGX H100 (4-GPU)
4 GPUs/system
4.1K

Specifications

Primary GPUNVIDIA H100 SXM5 80GB
GPU Count16.4K
SectorPrivate

Additional Information

"With high-performance local storage, high-performance computing (HPC) storage, cluster networking and memory, bare metal instances can be part of an OCI Supercluster that can scale to tens of thousands of NVIDIA H100 GPUs." "An OCI Supercluster with H100 GPUs can scale up to 16,384 GPUs"

Unclear when this was first operational. In November 2023, they state that OCI superclusters "can scale up to" 16k H100s. However, it's unclear whether they had actually built a 16k cluster yet, or were just accepting orders. Musk's xAI was said to be renting 16k of these GPUs in May 2024, signaling that at least one Oracle cluster of this size was operational by then. I mark the first operational date as April 2024 to be slightly conservative. Also, it's very likely that Oracle has several of these 16k H100 clusters