24 HGX nodes

• Nodes alphagpu01–18 — 8× H100 80GB GPUs per node
• Nodes alphagpu19–24 — 8× H200 141GB GPUs per node

• 10× 400Gb/s ConnectX-7 NIC cards per node (8 for InfiniBand, 2 for Ethernet)
• 30TB NVMe caching space per node
• 2TB of system memory per node

4 RTX PRO 6000 Blackwell nodes

• Nodes alphagpu51–54 — 8× RTX PRO 6000 Blackwell ~96GB GPUs per node (32 total)
• Part of the same general-access alpha partition as the H100/H200 nodes, reserved for single-GPU workloads — see the Job Submission and QoS Overview article for routing behavior

1 CPU node

• Node alphacpu01 — 2× Intel Xeon Platinum 8568Y+ (96 cores total, x86_64), ~1TB RAM
• Part of the dedicated cpu partition for CPU-only workloads

60 Grace CPU nodes

• Nodes betagg01–60 — 2× NVIDIA Grace (Neoverse-V2) sockets per node, 144 cores total, ARM (aarch64), ~480GB RAM per node
• CPU-only, part of the grace partition

1 Grace Hopper node

• Node alphagh01 — NVIDIA GH200 Grace Hopper Superchip: 1× Grace CPU (72 cores, ~573GB system memory) + 1× Hopper GPU (~96GB HBM3e)
• Part of the grace partition, alongside the 60 Grace CPU nodes above

Non-blocking NDR fabric, cabled for rail configuration

• 8 network switches, 96 optical connections

2 service nodes

• 1 login nodes
• 1 cluster management nodes (NVIDIA Base Command, licensed for all gear)

2 data transfer nodes

~10PB of DDN storage (Lustre/EXAScaler)

• Working/scratch tier for active training checkpoints

~20PB of VAST storage

• Home directories (100GB/user hard quota) + project directories