# Per-Layer OI/CF Roofline — Llama-3.1-8B on H100 (Section 4.4)

Per-component operational intensity (OI) and capacity footprint (CF) roofline
decomposition for a Llama-3.1-8B decoder layer on a single H100-SXM5-80GB
(bf16, TP=1). Backs Section 4.4 "Layer-Level Roofline Decomposition" and
Figure `per_layer_oi_cf_8b_h100.pdf`.

## Files

- `analytical/` — Per-component OI computed from model dimensions.
  B=1 prefill (C=1, memory-limited) and B=80 prefill (C=80, compute-saturated).
- `ncu/` — NCU v2024.3.2 kernel traces (`--set full`) on 8 decoder layers
  (indices 1–8) at bs=1 prefill. Clean run, 52 us total GPU time across
  8 unique kernel types. Reproduce with:
  ```
  CUDA_VISIBLE_DEVICES=0 ncu --set full --csv --profile-from-start no \
    --target-processes all -o output \
    python3 scripts/roofline/_ncu_target_simple.py
  ```
- `per_layer_oi_cf_8b_h100.pdf` — Two-panel figure: (a) per-component OI on
  H100 roofline, (b) capacity footprint sweep C=1..80.
- `per_layer_table.tex` — LaTeX table classifying each component as
  compute-bound (OI > 295 FLOP/byte) or bandwidth-bound.

## Key findings

| Component | Wt (MB) | OI (C=1) | OI (C=80) | Bound |
|-----------|---------|----------|-----------|-------|
| GEMM projections (Q,K,V,O,Gate,Up,Down) | 436 | 315–441 | 803–2956 | COMPUTE |
| RMSNorm (in/out), RoPE, SiLU, Gate×Up | — | 0–1 | 0–1 | BANDWIDTH |
| Flash attention | — | 64 | 64 | BANDWIDTH |
| **Layer total** | **436** | **330** | **878** | **COMPUTE** |

NCU validation: at C=1, GEMM = 10.8% of GPU time, memory ops = 65%,
confirming the layer is barely compute-bound (OI=330 vs ridge=295).

System capacity-bound at 32K context: C ≥ 16 (KV cache + weights > 80 GB HBM).

## Citation

AgentPerfBench: A Benchmarking and Evaluation Suite for Inference Performance
of Agentic LLMs. NeurIPS 2026.