Answers are grounded in this paper and cite their sources.
Framework folklore says that automatic mixed precision buys about 1.3x, that torch.compile buys about 1.1x, that gradient checkpointing trades a large block of memory for a moderate slowdown, and that a hand-written training loop and a high-level framework running the same recipe should land in the same place. This study froze those priors as seven hypotheses with verdict rules in a pre-registration (PREREG.md, the root commit of the repository, before any training run) and measured them on one fixed task: ResNet-18 on EuroSAT land-use classification (27,000 images, 64x64, 10 classes), trained under one frozen recipe on a consumer laptop GPU (NVIDIA GeForce RTX 5050 Laptop GPU, 8 GB) with two CPU arms on a 4-core virtual machine. Six of the seven hypotheses are not supported. The hand loop is slower than the PyTorch Lightning baseline (throughput ratio 0.687, 95% CI [0.682, 0.692]) and also more accurate (paired test difference +3.43 percentage points, CI [+2.80, +4.09]), so the equivalence hypothesis fails in both directions. AMP reaches a ratio of 1.127 (CI [1.121, 1.134]) against the 1.30 prior, with accuracy preserved and peak memory down 27.8 percent. torch.compile reaches 1.027 (CI [1.016, 1.039]) against the 1.10 prior. Gradient checkpointing saves 14.8 percent of peak memory at batch 128 against a 30 percent criterion, and costs almost nothing (ratio 0.966). CPU DDP with gloo tracks the single-process loss trajectory (mean absolute difference 0.017 nats) but misses the endpoint accuracy window by 9.59 percentage points. The one supported hypothesis is the profiler hypothesis: at the baseline loader configuration, data wait is 53.7 percent of step time, and the registered optimization (four workers, pinned memory, persistent workers) cuts it to 1.0 percent and mean step time from 81.3 ms to 29.5 ms. One fact organises the table: on this machine and task, training is bound by data loading, not by GPU compute, and priors imported from compute-bound workloads did not survive contact with a 64x64 dataset. The pre-registration is what makes that statement evidence instead of anecdote.
No takes yet. Share an insight, caveat, or question.
Jesun Ahmad Ushno (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: