Tillicum System Architecture
Tillicum is the University of Washington’s next-generation AI-accelerated research computing platform, purpose-built for the most demanding scientific workflows in machine learning, data science, and simulation. Designed around cutting-edge NVIDIA H200 GPUs with ultra-fast NVLink 4.0 interconnect and a 400 Gbps InfiniBand network, Tillicum enables researchers to move beyond the limits of traditional HPC with transformative performance and efficiency.
Key Features
- 24 Dell XE9680 servers
- GPU Acceleration: 192 NVIDIA H200 GPUs with 141 GB memory and ultra-high-bandwidth NVLink 4.0
- High-Speed Networking: 400 Gbps NDR InfiniBand for low-latency GPU communication
- High-Performance Storage: Shared 3 PB flash storage optimized for active, high-throughput computing
- Common Datasets: Curated datasets preloaded or available upon request for AI and data science workloads
- Flexible Scheduling: Support for interactive, batch, and multi-node workloads
- Container Support: Apptainer and Docker-compatible environments
- Research Software Support: Optimized deep learning frameworks (e.g., PyTorch, TensorFlow), MPI, CUDA, and more
| Component | Details |
|---|---|
| Cluster | tillicum |
| Service | Deployed August 2025, Early Access August 28-October 6, 2025, General Availability October 15 2025 - |
| Operating System | Rocky 9 |
| Compute | 1,536 CPU cores [Intel Emerald Rapids] |
| Accelerators | 192 GPU cards (141 GB memory per GPU) [NVIDIA Hopper – H200 SXM] |
| GPU Interconnect | NVLink™ 4.0 (900 GB/s GPU-to-GPU bandwidth) |
| Interconnect | 400 Gbps NDR InfiniBand |
| Storage | ~3 PB high-performance flash storage for active computing |
| Namesake | Tillicum is a word in Chinook Jargon, meaning “the people.” |