The single-card ceiling
The single-card ceiling¶
| rows | train() wall | peak device mem | throughput | train r2 (1M sample) |
|---|---|---|---|---|
| 300M | 328s | 42.2 GiB | 54.8M rows/s | 0.8305 |
| 400M | 414s | 56.0 GiB | 57.9M rows/s | 0.8303 |
| 450M | 465s | 63.0 GiB | 58.0M rows/s | 0.8340 |
| 500M | 513s | 69.9 GiB | 58.5M rows/s | 0.8329 |
Source: single-card-ceiling-2026-07.jsonl. Single-pod ladder (2026-07-18, NVIDIA A100-SXM4-80GB): a 500M x 100 float32 matrix trains end to end on one 80GB card at 69.9 GiB peak, 60 rounds in 8.5 minutes, with the device-resident objective keeping the fit loop bus-free. Evidence: benchmarks/single-card-ceiling-2026-07.md.