Arizona State University's Research Computing team describes, in a PEARC '26 paper, how it provisioned a donation of 229 Intel Gaudi2 servers (more than 1,600 accelerators) received in fall 2025 and installed late that year. Before the donation, the Sol supercomputer offered roughly 300 NVIDIA A100 GPUs for AI work. The goals were to widen generative-AI access across the university beyond traditional HPC users and to offload suitable work, such as PyTorch jobs and ESMFold protein folding, from scarce NVIDIA GPUs.
Two modes. About 200 nodes in a secondary data center run bare-metal Kubernetes as a service platform, while 15 nodes in the primary data center sit next to Sol's shared scratch storage; ten of those run batch jobs under Warewulf and Slurm. The two sites are joined by a 400 Gbps link, too slow for I/O-heavy batch jobs at the remote site, which is why it runs services instead. Each node has dual Xeon Platinum 8368 CPUs, 512 GB RAM, a 7 TiB NVMe drive used as a local model-weight cache, and eight HPUs with 96 GB HBM2e each, supporting context windows up to 256K tokens. Power limited racks to three servers each, since a node draws up to about 5 kW. Nodes network-boot a single golden image through Warewulf.
Service stack. HAProxy terminates TLS and routes traffic. LiteLLM is the API gateway for key validation and token tracking, AIBrix (built on Envoy Gateway) routes and load-balances, and the Habana Gaudi2 fork of vLLM serves each model as its own Kubernetes service. Users request API keys through a self-service portal tied to campus single sign-on. Prometheus and Grafana monitor the cluster, Langfuse handles token accounting, and namespaces and RBAC enforce least privilege. Access paths include OpenWebUI with SearXNG web search, VS Code and OpenCode integration, Jupyter AI, visual agent-workflow tools and programmatic API calls; ASU's CreateAI builder also exposes AIR-hosted models such as Gemma 4.
Results and limits. In the first half of 2026, over 258 users obtained API keys and sent over 47 billion tokens to open models on the Kubernetes side, and the ten Slurm nodes logged more than 1.5 million hours for over 100 users in 46 teams, including four AI courses. The team found Gaudi2 strong on dense, regular workloads but slowed by graph recompilation when input shapes vary, so batching consistent shapes helps. A port of the Evo2 genomic model only reached proof of concept, because Gaudi2 lacks complex-number support for its FFT layers. The six 400 Gb interfaces per node are unused for now; the plan is five-node "supernodes" of 40 HPUs for larger models.