{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,6,30]],"date-time":"2026-06-30T19:03:53Z","timestamp":1782846233902,"version":"3.54.5"},"reference-count":0,"publisher":"ECMS","content-domain":{"domain":[],"crossmark-restriction":false},"short-container-title":[],"published-print":{"date-parts":[[2026,6,23]]},"abstract":"<jats:p>We present a comparative evaluation of three transformer-based time series models\u2014iTransformer, PatchTST, and FEDformer\u2014across two deep learning frameworks: the original PyTorch implementations and newly developed ports to Apple\u2019s MLX framework.\n\nExperiments are conducted on the ETTm1 electricity transformer dataset, covering long-term forecasting and anomaly detection tasks. Using identical hyperparameters and a consistent 80\/20 chronological train\u2013validation split, we measure training throughput, per-epoch wall-clock time, memory consumption, and predictive accuracy (RMSE, MAE for forecasting; precision, recall, F1, AUC for anomaly detection).\n\nResults indicate that MLX achieves substantially higher training throughput than PyTorch on the M4 Pro GPU, with speedups ranging from 1.02\u00d7 to 3.12\u00d7 across model architectures and batch sizes. Forecasting accuracy remains comparable between frameworks, with differences in RMSE below 6%.\n\nFor anomaly detection, MLX attains higher F1 and AUC scores under identical configurations. These findings suggest that MLX is a viable alternative for time series model development on Apple Silicon, offering competitive accuracy with improved training efficiency.\n\nWe attribute the observed speedups to architectural differences in memory handling and execution models, highlighting how framework design interacts with Apple Silicon\u2019s unified memory architecture.<\/jats:p>","DOI":"10.7148\/2026-0638","type":"proceedings-article","created":{"date-parts":[[2026,6,30]],"date-time":"2026-06-30T18:08:43Z","timestamp":1782842923000},"page":"638-644","source":"Crossref","is-referenced-by-count":0,"title":["Benchmarking transformer-based time series models: pytorch vs mlx on apple silicon"],"prefix":"10.7148","author":[{"given":"Kamil","family":"Dziedzic","sequence":"first","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Mateusz","family":"Nytko","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Filip","family":"Kruzel","sequence":"additional","affiliation":[],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"4144","published-online":{"date-parts":[[2026,6,23]]},"event":{"name":"40th ECMS International Conference on Modelling and Simulation"},"container-title":["ECMS 2026 Proceedings edited by Filippo Sanfilippo, Florenc Demrozi, Fabio Sgarbossa, Mohammad Poursina"],"original-title":[],"deposited":{"date-parts":[[2026,6,30]],"date-time":"2026-06-30T18:08:44Z","timestamp":1782842924000},"score":1,"resource":{"primary":{"URL":"http:\/\/www.scs-europe.net\/dlib\/2026\/ecms2026acceptedpapers\/0638_dis_ecms2026_0035.pdf"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,6,23]]},"references-count":0,"URL":"https:\/\/doi.org\/10.7148\/2026-0638","relation":{},"subject":[],"published":{"date-parts":[[2026,6,23]]}}}