PyTorch Speeds Up AI Math on Phones, Macs and ARM Chips
Quicker AI on your phone — and a little less battery drain, too.
PyTorch is the free software engine under the hood of a huge share of AI tools, from image generators to chatbots to the smart features in your phone's photo app. This update fixes a bottleneck in how PyTorch does math on ARM chips — the processors inside iPhones, Macs, most Android phones, Raspberry Pis and many cloud servers. On those chips, certain calculations were being done one number at a time. Now they can be done several at once, which is like a supermarket checkout scanning four items in a single beep instead of one by one.
The fix uses Sleef, a specialised math library for calculating things like logarithms and angles. PyTorch already used it on Intel and AMD chips, but ARM machines never got it, despite years of the same code being present. Two separate barriers blocked it: a list of approved math functions that only named Intel-style ones, and a check that looked for an Intel-only chip feature that simply never appears on ARM. This change registers the ARM versions and removes that check. Engineers also had to teach the code to ask for four numbers at a time rather than eight, since ARM's built-in shortcut tops out there.
So what does that mean for you? AI features that run on your own device — background blur in video calls, on-phone photo editing, voice transcription, translation — lean on this kind of math. Faster math means snappier responses and less battery used per task. It also helps cloud servers running on ARM chips, which are cheaper to rent, so some companies may pass those savings along. As a bonus, the same fix slightly improves certain calculations on regular Intel and AMD machines too.
The catch: this is plumbing, not a new product. You won't see a new button or a new feature, and the speed-up applies only to specific math operations, not everything PyTorch does. The gains will also be modest unless a developer updates PyTorch and rebuilds their app. Nothing here changes what data you share or who can see it — it's simply the engine running a bit more efficiently.
- PyTorch, the free engine behind most AI apps, just made certain math run several numbers at once on ARM chips instead of one at a time.
- It uses a math library called Sleef, which until now only worked on Intel and AMD chips — not the ARM chips in iPhones, Macs and Android phones.
- Expect slightly faster on-device AI features and a little less battery drain, once app makers update to the newest PyTorch.
Why It Matters
Snappier AI features on your phone or Mac, plus a bit less battery used per task.