HPC Cluster Planner
Plan an HPC cluster from its node design: theoretical peak (Rpeak) from CPUs and GPUs, an estimated sustained figure from an efficiency you set, plus power, heat and the racks needed by space and by power.
Results
- CPU peak (FP64)
- 144.2 TFLOPStheoretical
- GPU peak (FP64, vendor)
- 0 TFLOPSVendor-reported
- Cluster Rpeak
- 144.2 TFLOPStheoretical
- Estimated sustained
- 100.9 TFLOPSestimated
- IT power
- 35.2 kWcalculated
- Heat load
- 1,20,107 BTU/hcalculated
- Racks needed1 by space, 2 by power
- 2calculated
- Peak per watt
- 4.1 GFLOPS/Wtheoretical
Power, not space, limits rack density. Consider higher-power racks or liquid cooling.
What it calculates
- CPU and GPU peak FP64
- Cluster Rpeak
- Estimated sustained performance
- Power and heat
- Racks by space and by power
- GFLOPS per watt
Worked example
Using the default inputs below; change any input above to recalculate.
| Compute nodes | 32 |
|---|---|
| CPUs per node | 2 |
| Cores per CPU | 64 |
| Sustained vector clock | 2.2 GHz |
| FP64 FLOPs per cycle per core | 16 (AVX2 2 FMA or AVX-512 1 FMA) |
| GPUs per node | 0 |
| GPU FP64 peak | NVIDIA H100 SXM5 80GB — 34 TFLOPS |
| Sustained efficiency (HPL-like) | 70 % |
| Power per node (max) | 1100 W |
| Rack units per node | 1 U |
| Usable U per rack | 38 U |
| Power limit per rack | 20 kW |
| CPU peak (FP64) | 144.2 TFLOPS | theoretical |
|---|---|---|
| GPU peak (FP64, vendor) | 0 TFLOPS | Vendor-reported |
| Cluster Rpeak | 144.2 TFLOPS | theoretical |
| Estimated sustained | 100.9 TFLOPS | estimated |
| IT power | 35.2 kW | calculated |
| Heat load | 1,20,107 BTU/h | calculated |
| Racks needed | 2 | calculated |
| Peak per watt | 4.1 GFLOPS/W | theoretical |
Methodology
CPU peak = nodes × CPUs × cores × sustained clock × FP64 FLOPs per cycle. GPU peak uses vendor-reported FP64 figures. Sustained performance applies your efficiency factor. Racks are the larger of the counts required by rack units and by the per-rack power limit.
Formulas
- Rpeak
- R = Σ_CPU (cores × f × FLOPs/cycle) + Σ_GPU peak_FP64
- Racks
- max(⌈nodes × U ÷ usable_U⌉, ⌈nodes ÷ ⌊rack_kW ÷ node_kW⌋⌉)
Assumptions
- FP64 vector peak is used for CPUs; GPU FP64 figures are non-tensor vendor values.
Limitations
- Efficiency is an input; real codes range widely.
- Switches, storage and management nodes are not included.