@TheAhmadOsman: The optimizations in MLX are horrid, DHH is right Apple doesn’t have a hardware problem they have a software problem
Summary
A tweet criticizes the optimizations in MLX, suggesting Apple has a software problem rather than a hardware issue, based on performance tests with M5 Ultra Macs.
View Cached Full Text
Cached at: 09/25/26, 02:40 PM
The optimizations in MLX are horrid, DHH is right Apple doesn’t have a hardware problem they have a software problem
Alex Ziskind (@digitalix): Faster chips, new bottleneck 🤯 Two clustered M5 Ultra Macs ran SLOWER than one: 35 tok/s vs 53. Which is strange, considering my experiments with M3 Ultra were the opposite.
Keep the GPUs awake and they jump to 58 tok/s.
Similar Articles
Mac Heads: Is there any point to MLX in September 2026?
A user questions the relevance of MLX on Apple M5 chips after llama.cpp updates for Metal improve GGUF model prefill performance, making MLX potentially redundant for certain tasks.
@TheAhmadOsman: To be clear, CUDA is still more mature than MLX if you’re comparing GPUs to the new M5 Ultra Mac Studio However, the ec…
The discussion compares CUDA's maturity to MLX in the context of Apple's M5 Ultra Mac Studio, stressing the need for competition to optimize hardware for local AI, which could enable Apple to surpass NVIDIA.
@awnihannun: Pretty remarkable what two M5 Ultras + MLX can do now
Awni Hannun 展示了两台 M5 Ultra 配合 MLX 所能达到的性能水平，凸显本地 Apple Silicon 推理能力的显著提升。
@ivanfioravanti: Apple M5 Max + MLX = raw power! Look at this demo I'm playing with "FasterLivePortrait-MLX" I started with MPS but resu…
The author demonstrates that migrating a LivePortrait implementation from MPS to Apple's MLX framework on an M5 Max chip results in significantly better performance and speed.
Apple M5 isn't making full use of its matmul cores yet
Apple M5 silicon supports INT8 activations for matrix multiplication, but inference backends like MLX and Llama.cpp currently use 16-bit; custom w8a8 kernels achieve up to 1.4x speedup on Gemma4 prefill tasks.