Skip to content

10_spec_sheet.md

Latest commit

 

History

History
40 lines (32 loc) · 2.64 KB

10_spec_sheet.md

File metadata and controls

40 lines (32 loc) · 2.64 KB
Raw

Torch Spec Sheet

The Torch cluster has 518 Intel "Xeon Platinum 8592+ 64C" CPUs, 29 NVIDIA H200 GPUs & 68 NVIDIA L40S GPUs connected together via Infiniband NDR400 interconnect. Further details on each kind of node is provided in the table below.

Torch System Resources

Type Nodes CPU Cores GPUs Memory (GB) CPUs per Node GPUs per Node Memory per Node (GB)
Standard Memory 186 23,808 N/A 95,232 128 N/A 512
Large memory 7 896 N/A 21,504 128 N/A 3,072
H200 GPU 29 3,712 232 59,392 128 8 2,048
L40S GPU 68 8,704 272 34,816 128 4 512
RTX Pro 6000 GPU 2 256 16 3,072 128 8 1,536
H100 GPU 15 1,440 60 23,040 96 4 1,536
A100 GPU 9 576 36 4,608 64 4 512
A100 GPU 34 2720 136 34,816 80 4 1,024
Login 4 512 N/A 1024 128 N/A 256
Data Transfer 2 64 N/A 512 32 N/A 256
Provisioning 4 320 N/A 1024 80 N/A 256
Scheduler 2 64 N/A 1024 32 N/A 512
Total N/A 43,072 752 280.1(TB) NA NA NA

Torch was tested in June 2025 using the LINPACK benchmark system, which is the basis for all HPC systems ranked on the Top500 list. It had a theoretical maximum performance of 12.25 PF/s thanks to its powerful GPU resources, of which LINPACK was able to use 10.79 PF/s, thus placing it at #133 on the listed.

Torch was recently ranked #40 on the Green 500 list, a global list of the most energy efficient supercomputers in the world thanks to its advanced liquid cooling system.

Torch Stakeholder Resources

Many GPUs on Torch are owned by stakeholder groups who have prioirty access to the resources. The breakdown of these GPUs into stakeholder groups is as follows:

Stakeholders H200 L40S H100 A100 RTX Pro 6000 Total
Tandon SoE 112 60 60 232
Courant (CS) 88 52 140
CDS 24 24 48
Chemistry 56 56
CILVR 32 32
Individual PIs 24 12 16 52
IT Public 24 208 232
Total 272 272 60 172 16 792