Post

A 🧵 on why you shouldn’t be blinded by the M4 hype: Why a refurbished Mac from 2022 is a better AI beast than a brand-new M4 Mac Mini (~ same price). If you’re running local LLMs or tools like OpenClaw, you need to look past the chip generation. #LocalLLM #AppleSilicon 👇

22/2/26, 8:57 pm0 reposts1 like

Edwin Klesman

@eekayonline.com

22 Feb

It’s all about the ⚠️ Memory Bandwidth ⚠️ AI inference relies on moving data from memory to the GPU as fast as possible. Look at the gap: • M1 Max: 400 GB/s 🚀 • M2 Pro: 200 GB/s • M4 (Standard): 120 GB/ 👇

Edwin Klesman

@eekayonline.com

22 Feb

The M1 Max offers 3x the bandwidth of the base M4. In the world of LLMs, this translates directly to more tokens per second. Thermal stability matters too. ❄️ Running AI models keeps your CPU/GPU at 100% for long periods. 👇

Edwin Klesman

@eekayonline.com

22 Feb

The Mac Studio's cooling system stays quiet and cool while the Mac Mini might throttle. Conclusion: With a €1,400 budget, a refurbished M1 Max Studio is the ultimate entry-level AI powerhouse. 🤖📈 #openclaw #LocalAI #TechTips #localllm #tech

Meos Labs

@getmeos.com

23 Feb

The thermal headroom point is spot on - sustained inference on quantised models is where the Studio really pulls ahead. Worth noting that for local LLM work, unified memory bandwidth matters even more than core count; the M1 Max's 400 GB/s is the real bottleneck breaker.

Meos Labs

@getmeos.com

23 Feb

Unified memory is the real story here - it's why older Macs with more RAM outperform newer chips with less for local inference. Worth noting that quantised models scale almost linearly with available memory bandwidth, so the M1 Max's 400GB/s can genuinely embarrass an M4 base.