Callosum ↗
Heterogeneous Orchestration
Cluster 2 explores how heterogeneous hardware orchestration can unlock new levels of inference efficiency for agentic AI workloads. By combining different accelerator architectures under a unified orchestration layer, the cluster tests whether intelligent workload distribution can optimise performance across mixed hardware environments.
Cluster specification
Live · Racks A4–A5, B4–B5
60
GPUs
15.5 TB
GPU memory
1,200
CPU cores
20.7 TB
System memory
B300 HGX × 3
Supermicro AS-8126GS-NB3RT · 8× B300 SXM 288GB · 2× EPYC 9575F · 2.3TB
MI350X × 3
Supermicro AS-8126GS-TNMR · 8× MI350X 288GB · 2× EPYC 9575F · 2.3TB
H200 NVL × 3
Supermicro SYS-422GL-FNR2 · 4× H200 NVL 141GB · 2× Xeon 6960P · 2.3TB
Network (B300)
Fabric: 8× 800G OSFP · FE: 2× 400G
Network (MI350X, H200)
Fabric: 8× 400G QSFP112 · FE: 2× 400G
Non-blocking leaf-spine fabric: 8× NVIDIA SN5610 (6 leaf + 2 spine) in rack B5, plus 2× SN2201 management switches.
Rack layout
