Design

Heterogeneous Cluster Designer

Compose a cluster from multiple GPU pools — mix types (H100 + L40S) or vendors (NVIDIA + AMD). Each pool is a labelled, tainted node group; workloads are steered to the right GPUs. The designer aggregates capacity, power, and cost across pools.

1NVIDIA
GPU
Role
Nodes · 4
GPUs/node · 8
2NVIDIA
GPU
Role
Nodes · 4
GPUs/node · 8
Cluster totals · 2 pools
Total GPUs

64

Aggregate FP16

43.3 PF

GPU memory

4,096 GB

Racks

4

~15 kW/rack

Facility power

56 kW

Est. CapEx

$1.25M

GPUs only

Est. energy/mo

$5K

Vendors

NVIDIA

Heterogeneous scheduling
  • Label and taint each pool by GPU type; target workloads with nodeSelector/affinity + matching tolerations.
  • Keep a single distributed-training job on one homogeneous pool — collectives (all-reduce) need matched GPUs.
  • Same-vendor mix: one GPU Operator manages all nodes; differentiate pools with the gpu.product node label.
Topology

Rendering diagram…

Bill of materialsEstimate
Pool 1 · H100 SXM5 GPU328/node × 4 · training
Pool 2 · L40S GPU328/node × 4 · inference
InfiniBand (RDMA)sharedRDMA
Parallel file system1 tiershared
Rack4~15 kW usable/rack
Kubernetes control plane3HA, shared across pools

Planning estimates from illustrative pricing — not quotes.