social-data-attribution-w2/docs/PACE_ICE_MAKERSPACE.md

# AI Makerspace GPU Resources (PACE ICE)

Source: PACE ICE Cluster Resources documentation (March 2026)

## GPU Nodes Reserved for CoE / AI Makerspace

| Quantity | CPU | Memory | GPU | Local NVMe |
|----------|-----|--------|-----|------------|
| 14 nodes | Dual Xeon Platinum 8462Y+ (64 cores, 2.8GHz) | 2048GB DDR5 | 8x H100 SXM5 80GB | 3x 3.84TB |
| 12 nodes | Dual Xeon Platinum 8462Y+ (64 cores, 2.8GHz) | 2048GB DDR5 | 8x H200 SXM5 142GB | 8x 3.84TB |

**Totals:** 112 H100 GPUs + 96 H200 GPUs = 208 GPUs

## Access

- Partition: `coe-gpu` (automatic for CoE/AI Makerspace users)
- Also accessible via general `ice-gpu` partition at lower priority
- No charge account needed (ICE is free)
- Jobs are auto-routed to eligible partitions

## Storage

- `~/scratch`: 300GB quota, 120-day purge
- Network storage limit: ~2TB usable
- Node-local `$TMPDIR`: 11.5TB (H100) or 30.7TB (H200) — cleaned at job end

## Network

- InfiniBand HDR100 (100 Gbps) between nodes
- Internet egress for R2 streaming: shared bandwidth