This is nowhere near a deep analysis. I own both a MacBook Pro M3 Max with 128 GB of unified memory and a system equipped with an RTX 5090 and 128 GB of RAM, so I can compare them directly under real-world workloads.

The RTX 5090, running the Basalt Engine (https://github.com/jesdga95/basalt), significantly outperforms the M3 Max across every metric I've tested. I'm seeing approximately **180 tokens/s in generation and 5,000 tokens/s in prefill**.

That second figure is arguably the most important metric for agentic coding harnesses, where large contexts, frequent tool calls, and repeated prompt processing can make prefill performance far more consequential than raw generation speed.

Any serious comparison needs to account for actual inference engines, prefill throughput, generation speed, context handling, and real-world agentic workloads. Comparing hardware specifications or isolated benchmarks without considering these factors misses the bigger picture.