Apple’s M4 Powers Fanless MacBook Air to 2.78 TFLOPS in New MLX Benchmark

Apple’s M4 chip hits its stride in a matrix‑multiplication test that pushes a fanless MacBook Air to 2.78 TFLOPS, the highest figure recorded in the benchmark series. The test, run with Apple’s MLX v0.28 (and later verified on v0.31.1) on Python 3.10.11, multiplies two 20 000 × 20 000 bfloat16 matrices, completing roughly 16 trillion floating‑point operations in 5.75 seconds. That translates to a real‑world throughput of 2.78 TFLOPS, according to the benchmark author, a Japanese‑language blogger who posted the results on lwgena for TinyAlg on March 19. The script, publicly available on GitHub Gist, can be reproduced on any Apple‑silicon Mac with a single “pip install mlx” command, and the author confirmed it still runs unchanged on Python 3.12.12.

The performance curve reveals how the fanless chassis copes with sustained load. While the 10 000 × 10 000 run peaks at 2.70 TFLOPS with a sub‑second latency, the 30 000 × 30 000 test drops to 2.50 TFLOPS as execution time climbs to 21 seconds. The author notes a gradual creep in runtime for the 20 000 and 30 000‑size matrices, suggesting the thin‑profile thermal spreader is beginning to throttle the silicon as heat accumulates. This mirrors observations from earlier fanless models: 9to5Mac’s review of the M3 MacBook Air highlighted that Apple’s passive cooling can sustain respectable CPU and GPU bursts but eventually yields to thermal limits under prolonged stress.

Memory constraints become the dominant bottleneck at the 40 000 × 40 000 benchmark. The 16 GB of unified RAM fills, forcing the system to swap heavily to the SSD; execution time spikes to 75 seconds and variance between runs widens dramatically. The author captured the surge in swap usage via Activity Monitor, confirming that the slowdown is not a CPU issue but a bandwidth one. This mirrors the architecture of the M2 Air, whose internals were dissected by Wccftech, showing a heatsink‑covered SoC but no active fan, reinforcing that Apple’s fanless designs rely on a delicate balance of memory bandwidth, silicon efficiency, and passive heat dissipation.

Comparing the M4’s raw math throughput to its predecessors underscores a steady climb in on‑device AI capability. The M3 Air, reviewed by 9to5Mac, delivered respectable performance for everyday tasks but was still positioned as a “low‑power” machine. The M4’s 2.78 TFLOPS figure, achieved without a fan, rivals entry‑level desktop GPUs and exceeds the M3’s reported benchmarks in similar matrix‑multiply workloads. Apple’s decision to ship the M4 in a fanless Air suggests confidence that the new generation’s neural‑engine and matrix‑multiply units can deliver AI workloads—such as local inference for LLMs—without overheating, a claim the benchmark author supports with empirical data.

The broader implication for developers is clear: Apple’s MLX framework, now at version 0.31.1, can tap the M4’s matrix engine directly, delivering teraflop‑scale performance on a portable laptop. For teams building on‑device models, the 2.78 TFLOPS result signals that the Air can handle sizable transformer layers or diffusion pipelines without resorting to cloud off‑load. As the AI industry pushes toward edge inference, Apple’s fanless Air may become a reference point for performance‑per‑watt, challenging the narrative that high‑end AI compute requires bulky, fan‑cooled hardware.