# MSC — Minimum Sufficient Compute Artifacts for *Is Compute Difficulty Architecture-Agnostic? Measuring and Distilling Per-Sample Minimum Sufficient Computation*. Generated 2026-08-06T03:56:37Z by msc_lib v1.0.0. ## Repositories - `Shanmuk4622/msc-cifar100` — everything, one folder per run ## What MSC is The smallest cost-normalised configuration at which a network's decision has *stably settled* to its full-compute decision, defined uniformly over depth, resolution and precision reduction. Stability means the decision agrees at that budget **and every larger one** — predictions under compute reduction are not monotone, so a naive minimum records an accident rather than a property. ## Compute grid - depth: exits at [0.2, 0.4, 0.6, 0.8, 1.0] of network depth - resolution: [16, 20, 24, 28, 32] px, measured natively AND via a downsample-upsample proxy (the proxy cost model is labelled idealised) - precision: ['int4', 'int6', 'int8', 'fp16', 'fp32'], simulated by fake quantisation; cost priced analytically as bits/32, never as measured latency - confidence thresholds: tau in [0.0, 0.1, 0.2, 0.3, 0.5] — all results are tau-curves ## Per-image table schema ``` sample_idx, label, pred_d1..d5 top1p_d1..d5 top2p_d1..d5 depth pred_rn1..rn5 top1p_rn1..rn5 top2p_rn1..rn5 resolution (native) pred_rp1..rp5 top1p_rp1..rp5 top2p_rp1..rp5 resolution (proxy) pred_q1..q5 top1p_q1..q5 top2p_q1..q5 precision msp, margin, entropy, ce_loss, el2n, forget_events, pred_depth sample_order_hash, run_id, split ``` Every table carries `sample_order_hash`. Tables whose hashes differ are not row-aligned and must not be correlated. ## Telemetry recorded per epoch Losses and accuracies; learning rate per group; gradient norm mean/max/p95; gradient-clip hit rate; weight norm; update-to-weight ratio; AMP scale; NaN/Inf batch count; epoch/train/eval time; dataload vs compute split; step-time p50/p90/p99; throughput; VRAM allocated/reserved/peak; GPU utilisation and temperature; CPU and RAM; free disk; energy in J/kWh and CO2 per epoch and cumulative. Plus raw power samples at 10 Hz, system samples at 1 Hz, and a downsampled per-step trace. ## Reproducibility - `config.yaml` frozen at run start, sha256-hashed, asserted on resume - checkpoints carry optimizer, scheduler, AMP scaler and all four RNG streams - 3 seeds per headline number, mean +/- std - every artifact mapped to a run_id in `paper/provenance.csv` - work split across accounts by a deterministic cost-balanced scheduler; each run records which worker produced it ## Atlas results | arch | seed | acc_pct | reference_accuracy | params_M | GFLOPs | |:---------------|-------:|----------:|---------------------:|-----------:|---------:| | convnext_femto | 1 | 62.67 | nan | 4.87 | 0.126 | | convnext_femto | 2 | 62.6 | nan | 4.87 | 0.126 | | convnext_femto | 3 | 61.84 | nan | 4.87 | 0.126 | | mixer_nano | 1 | 60.23 | nan | 2.5 | 0.343 | | mixer_nano | 2 | 60.32 | nan | 2.5 | 0.343 | | mixer_nano | 3 | 60.72 | nan | 2.5 | 0.343 | | mobilenetv2 | 1 | 70.1 | 64.6 | 2.35 | 0.183 | | mobilenetv2 | 2 | 69.89 | 64.6 | 2.35 | 0.183 | | mobilenetv2 | 3 | 70.31 | 64.6 | 2.35 | 0.183 | | shufflenetv2 | 1 | 71.93 | 70.5 | 1.36 | 0.092 | | shufflenetv2 | 2 | 71.54 | 70.5 | 1.36 | 0.092 | | shufflenetv2 | 3 | 71.81 | 70.5 | 1.36 | 0.092 | | resnet110 | 1 | 74.31 | 74.31 | 1.74 | 0.511 | | resnet110 | 2 | 74.57 | 74.31 | 1.74 | 0.511 | | resnet110 | 3 | 74.26 | 74.31 | 1.74 | 0.511 | | resnet20 | 1 | 70.25 | 69.06 | 0.28 | 0.082 | | resnet20 | 2 | 70.36 | 69.06 | 0.28 | 0.082 | | resnet20 | 3 | 69.78 | 69.06 | 0.28 | 0.082 | | resnet32x4 | 1 | 79.59 | 79.42 | 7.42 | 2.149 | | resnet32x4 | 1 | 79.54 | 79.42 | 7.42 | 2.149 | | resnet32x4 | 2 | 79.63 | 79.42 | 7.42 | 2.149 | | resnet32x4 | 2 | 80.03 | 79.42 | 7.42 | 2.149 | | resnet32x4 | 3 | 79.46 | 79.42 | 7.42 | 2.149 | | resnet56 | 1 | 73.88 | 72.34 | 0.86 | 0.254 | | resnet56 | 2 | 73.35 | 72.34 | 0.86 | 0.254 | | resnet56 | 3 | 73.85 | 72.34 | 0.86 | 0.254 | | resnet8x4 | 1 | 73.35 | 72.5 | 1.22 | 0.334 | | resnet8x4 | 2 | 73.39 | 72.5 | 1.22 | 0.334 | | resnet8x4 | 3 | 73.04 | 72.5 | 1.22 | 0.334 | | vgg13 | 1 | 75.7 | 74.64 | 9.46 | 0.458 | | vgg13 | 2 | 75.65 | 74.64 | 9.46 | 0.458 | | vgg13 | 3 | 75.75 | 74.64 | 9.46 | 0.458 | | vgg8 | 1 | 71.56 | 70.36 | 3.96 | 0.136 | | vgg8 | 2 | 71.73 | 70.36 | 3.96 | 0.136 | | vgg8 | 3 | 71.61 | 70.36 | 3.96 | 0.136 | | vit_tiny | 1 | 59.33 | nan | 5.38 | 0.694 | | vit_tiny | 2 | 59.13 | nan | 5.38 | 0.694 | | vit_tiny | 3 | 60.56 | nan | 5.38 | 0.694 | | wrn_16_2 | 1 | 73.64 | 73.26 | 0.7 | 0.203 | | wrn_16_2 | 2 | 73.79 | 73.26 | 0.7 | 0.203 | | wrn_16_2 | 3 | 74.27 | 73.26 | 0.7 | 0.203 | | wrn_40_1 | 1 | 72.41 | 71.98 | 0.57 | 0.168 | | wrn_40_1 | 2 | 72.54 | 71.98 | 0.57 | 0.168 | | wrn_40_1 | 3 | 72.28 | 71.98 | 0.57 | 0.168 | | wrn_40_2 | 1 | 76.89 | 75.61 | 2.26 | 0.658 | | wrn_40_2 | 1 | 76.06 | 75.61 | 2.26 | 0.658 | | wrn_40_2 | 2 | 76.72 | 75.61 | 2.26 | 0.658 | | wrn_40_2 | 2 | 76.62 | 75.61 | 2.26 | 0.658 | | wrn_40_2 | 3 | 76.16 | 75.61 | 2.26 | 0.658 | ## Limitations - Per-image routing gives **no wall-clock speedup under batched inference** unless the batch is split by route. The deployment claim is scoped to batch-1 / edge / streaming. - INT4 and INT6 are simulated; no T4 kernel exists to time them. - The resolution proxy runs at 32 px; its cost is an idealised model. - Risk control is calibrated at epsilon=0.03 on a 5,000-image holdout, because epsilon=0.01 would need ~14,979 calibration images and the CIFAR-100 test set has 10,000. - Energy is measurement methodology, not a contribution. - T4-only hardware; CIFAR-100 scale.