Build the Pod
Spec a GPU pod that trains a 13B model in two weeks, within budget and the facility's power envelope.
Briefingclick to collapse
Your lab won funding to pretrain a 13B model on 300B tokens. Constraints from finance and facilities:
- Capex ≤ $2.0M (GPUs, hosts, network)
- Facility power ≤ 60 kW, including cooling (PUE 1.3)
- Training time ≤ 14 days
Pick the GPU, node count and network fabric. The estimate searches parallelism layouts (TP/PP/ZeRO/micro-batch) for each design and reports the best one that fits in memory. Lower energy is better. It's part of your score.
Your design8 GPUs per node
Capex✓ $0.74M/ $2.00M
Facility power✓ 23 kW/ 60 kW
Time to train✕ 47.3 d/ 14.0 d
Energy to train
26.0 MWh
Carbon
9.9 t
Best layout
tp4 pp1 zero1 mb4
MFU / bottleneck
57% · compute
Objectives2/3 met
- Capex ≤ $2.0M ✓
- Facility power ≤ 60 kW ✓
- Trains 13B × 300B tokens in ≤ 14 days
−33 pts each