Mac Studio vs AMD AI PC: Don't Buy Before Watching This

Описание: Apple's new Mac Studio ships with 1.2 terabytes a second of memory bandwidth, and on the models people actually run, that buys about 25 percent.

Join Membership for Early Access -    / @kaiexplainsyt  

Every comparison between the Mac Studio and AMD's Ryzen AI Halo turns on a bandwidth ratio of roughly 2.4 to 1. That ratio is real, and it measures decode: the model writing one word at a time against the memory bus. It does not measure prefill, the pass that reads your prompt before the first word appears, which runs on GPU compute instead. One developer timed both on an M1 Max and found that at 8,500 tokens of context, prefill was 94 percent of the total response time, and that Apple's own MLX framework lost the wall clock to a community GGUF build it beat two to one on the counter.

Tom's Hardware measured the Mac's lead falling from 2.26x on dense Gemma 4 12B to 1.6x on gpt-oss-120b to 25 percent on Qwen 3.6 35B-A3B, because a mixture of experts fires 5.1 billion of 116.8 billion parameters per token and gives the faster bus less to carry. On prompt processing the M4 Max was level with the AMD box, and in places behind it.

The deciding variable turns out to be software. The same Radeon 8060S delivered 5.1 teraflops without hipBLASLt and 36.9 with it. Four configuration changes are worth 14 to 20 percent of prefill on the AMD box. A free MIT repo measured 2.23x decode and 3.05x prefill over llama.cpp's HIP backend on the same silicon, cutting a 16,000-token job from 147 seconds to 58. And the cheapest way in is a trap: owners report the GMKtec EVO-X2 capped at around 117 gigs a second against the platform's 215, with ROCm broken on several boards.

WHAT THIS VIDEO COVERS

Why tokens per second measures the half of the job you don't wait on
How mixture-of-experts models erased most of Apple's bandwidth advantage
The same AMD chip running at 9 percent and 60 percent of its own peak
Four config changes worth 14 to 20 percent of prompt processing
Which Strix Halo boxes ship broken, and what the working ones cost
The buyer for whom the Mac Studio is still decisively correct

Timestamp:

00:00 Mac Studio M5 vs Ryzen AI Halo
03:10 Prefill vs Decode: The Hidden Bottleneck
05:08 Why MoE Models Change the Comparison
09:25 AMD's Software Advantage
10:02 What a Serious Local AI Workflow Looks Like
12:27 The Problem With Cheap Ryzen AI Halo PCs
15:50 Which Machine Should You Buy?

SOURCES

All Links On notion - https://app.notion.com/p/Links-572b04...

#minipc #localai #llm #inference