Compute ask
The pilot runs on our own single 16 GB GPU. Grant compute is for the scaled phase (M2–M3). These are planning estimates, not measurements. The pilot will replace the assumed throughput with measured numbers, and the request will be updated.
Estimation method
Training FLOPs ≈ 6 × N × T (N = 8×109 parameters, T = training tokens). This is the standard dense-transformer approximation, and it is conservative for LoRA. GPU-hours = FLOPs ÷ (peak bf16 throughput × utilisation). We assume an H100 at about 989 TFLOP/s dense bf16 and 35% utilisation, which gives about 3.5×1014 FLOP/s effective.
| Item | Tokens | FLOPs | H100 GPU-hours |
|---|---|---|---|
| Norwegian adaptation (LoRA), 1 run | 2×109 tokens | 9.6×1019 | ≈ 77 |
| SFT: 2 bases × 2 ablations × 3 seeds = 12 runs | 1×108 tokens each | 4.8×1018 each | ≈ 46 total |
| Evaluation: ~30 checkpoints × NorEval + EuroEval + BFCL + NO-FC | inference | — | ≈ 60 (assumed 2 h/checkpoint) |
| Subtotal | ≈ 183 | ||
| + 50% buffer (failed runs, hyper-parameter search) | ≈ 275 H100 GPU-hours |
The same work on our single 16 GB card would take months, and some runs (longer sequences, multiple seeds in parallel) don't fit at all. That's why the scaled phase needs grant compute. On TPUs the equivalent is a v4/v5e slice for a few weeks.
Programmes we are applying to
- Google TPU Research Cloud
- Lambda Research Grant
- Hugging Face community GPU grant (public demo Space)
Any support received will be acknowledged by name in the model cards, repository and report, together with how much compute was used.