LocalLLaMA debates Mac, AMD, RTX rigs

This is the kind of discussion that matters more than spec-sheet hype, because long-context chat exposes where local inference actually feels slow.

• –The thread frames prompt ingestion as the bottleneck, which lines up with broader community benchmarking that shows long chats punish weaker prompt-processing throughput fast.
• –AMD’s AI Max+ 395 looks attractive for large-model fit and respectable generation speed, but reported results depend heavily on backend and driver maturity.
• –Nvidia desktop GPUs still appear to hold the edge on prompt processing, especially for long contexts, even if unified-memory systems are easier for loading bigger models.
• –Apple’s unified-memory machines remain convenient and quiet for remote local inference, but they are often judged less favorably when prompt latency becomes the main metric.
• –For AI developers, this is fundamentally an infrastructure choice about memory capacity, bandwidth, software stack quality, and interactive feel—not just peak tok/s.