Apple M4 neural engine 35T ops/sec claim

The claim making the rounds is directionally right, but the number is off. Apple’s official M4 spec says the Neural Engine can reach 38 trillion operations per second, not 35. Apple introduced that figure with the M4 launch in May 2024, first in iPad Pro, and framed it as the company’s fastest Neural Engine yet. ### So what is the Neural Engine claim, exactly? It’s a peak throughput number for Apple’s dedicated AI block — basically the part of the chip meant to chew through matrix-heavy inference tasks without burning through battery the way a CPU would. Apple’s wording is “up to 38 trillion operations per second,” which also tells you this is a ceiling under specific conditions, not a promise that every model on every app will hit anything close to that. (apple.com) ### Why does the 35T vs 38T detail matter? Because these chip claims get repeated fast, and small errors turn into fake precision. “Roughly 35T” sounds harmless, but Apple did publish a number, and it was 38T. If you’re trying to compare Apple silicon with Qualcomm, Intel, AMD, or Microsoft’s Copilot+ PC messaging, you want the actual spec before building a bigger argument on top of it. ### Does TOPS tell you real model speed? (apple.com) Only partly. TOPS is like quoting a car’s horsepower without saying the weight, tires, road, or transmission. Real inference speed depends on model size, quantization, memory bandwidth, prompt length, software stack, and whether the workload lands on the Neural Engine, GPU, or CPU. Apple itself pitches M4 as a package — Neural Engine, ML accelerators in the CPU, GPU, and faster memory bandwidth together — not as one magic block doing everything alone. ### Where does Apple’s real edge come from? From integration. Apple built MLX as an Apple-silicon-native machine learning framework, and the project has turned into a real developer on-ramp for local inference and experimentation on Macs. MLX is explicitly optimized for Apple silicon, and Apple’s own open-source pages keep stressing the same thing — efficient ML on unified-memory hardware. That matters more than a headline TOPS number because it lowers the friction between “the chip can do AI” and “developers actually ship AI on it.” (apple.com) ### What about the “2–3× faster than cloud” line? Treat that as a use-case claim, not a universal benchmark. For simple or medium tasks, on-device can absolutely feel faster because there’s no round trip to a server, no queueing, and no network jitter. But Apple’s own Apple Intelligence design makes clear that some requests still go off-device to Private Cloud Compute when they need larger models or more reasoning depth. So the honest version is: local inference can beat cloud on latency-sensitive tasks, but not because the M4 somehow outmuscles a datacenter. (github.com) It wins by being right there. ### Is the “1.5 billion devices” point right? That number is stale. Apple has been saying its installed base of active devices is at an all-time high, and by 2025 that base was already well beyond the old 1.5 billion talking point from earlier years. Apple didn’t restate an exact installed-base number in the releases surfaced here, but the broad idea is still valid — Apple ships AI features into an enormous live ecosystem. (security.apple.com) ### So what should you take away? The clean version is simple. Apple’s official M4 Neural Engine claim is 38 TOPS. The more important story is that Apple is building a full edge-AI stack around that silicon — local models, MLX, unified memory, and a private cloud fallback when local isn’t enough. That’s why the M4 matters. Not because one viral number was huge, but because Apple made the whole pipeline shippable. (apple.com)