Skip to content

System Architecture

All Ferranti nodes run Rocky 8.8 and resource allocation is managed using Slurm (resource manager). Global storage is provided by a Weka file system. Inter-node communication uses a InfiniBand 400G non-blocking network.

The system is composed of 2 login nodes, 2 CPU-only nodes, and 15 H100 nodes, housed in 10 air-cooled racks.

Login Nodes

Ferranti's two login nodes have the following configuration:

Feature Login Node
CPUs: 2 x Intel Xeon Gold 6430, 32 cores
RAM: 1024GB DDR5-4800 ECC RAM
Local Storage: 700GB NVMe

CPU Nodes

The 2 CPU-only compute nodes have the following hardware:

Feature Specifications
CPUs: 2 x AMD EPYC 9654(96 cores, 2.4 GHz, 384 MB L3 Cache)
RAM: 2304GB DDR5-4800
Local Storage: 50TB
Theoretical Peak Performance: ~

GPU Compute Nodes

Feature Specifications Specifications
Total Nodes: 5 10
Accelerators: 8 Nvidia H100 / node 8 Nvidia H100 / node
Accelerator connect: SXM5 SXM5
GPU Memory: 80GB HBM3 / card (bandwidth: 3.35TB/s) 80GB HBM3 / card (bandwidth: 3.35TB/s)
CPUs: 96 cores: 2 x Intel Xeon 8468, 48 cores/die, 2.1 GHz 192 cores: 2 x AMD Genoa 9654, 96 cores/die, 2.4 Ghz
RAM: 2048GB DDR5-4800 2304GB DDR5-4800
NVIDIA Tensor Cores: 528 / card 528 / card
FP32 Cores: 16896 / card 16896 / card
FP64 Cores: 8448 / card 8448 / card
Local Storage: 24TB NVMe 60TB NVMe
HCA: 6x NVidia Mellanox ConnectX-7 NDR400 6x NVidia Mellanox ConnectX-7 NDR400
Theoretical Peak Performance: (FP16) 15832 TFLOPS ; (TF32) 7912 TFLOPS ; (FP32) 536 TFLOPS ; (FP64) 272 TFLOPS (FP16) 15832 TFLOPS ; (TF32) 7912 TFLOPS ; (FP32) 536 TFLOPS ; (FP64) 272 TFLOPS

Ferranti Interconnect

Ferranti has a fat tree InfiniBand interconnect topology with the following composition:

Feature Specifications
Number of core switches: 4
Number of edge switches: 7
Interconnect topology and type: NDR InfiniBand Fat Tree, non-blocking
Blocking factor: 1:1

Last update: October 18, 2024
Created: June 21, 2024