NVIDIA RTX 3090 vs Apple M3 Pro (36GB)

Analysis

The RTX 3090 and Apple's M3 Pro (36GB configuration) sit on opposite ends of the local-LLM hardware spectrum: a five-year-old discrete desktop GPU with genuine used-market NVLink upside, against a laptop-class unified-memory chip with more total capacity but a small fraction of the bandwidth. This comparison is really a bandwidth-versus-capacity tradeoff, not a straightforward winner.

The M3 Pro's 36GB of unified memory actually edges out the RTX 3090's 24GB, and that shows up directly in this site's tracked model list: 54 of the 99 models this site tracks fit natively on the M3 Pro at 8k context versus 53 on the RTX 3090, one more, purely from the extra 12GB of capacity. Qwen 3.6 27B is the clearest example: it fits the M3 Pro at the higher-quality Q6_K quant (25.43 GB, inside the chip's roughly 28 GB usable budget once macOS's reservation is subtracted) while the RTX 3090's smaller usable VRAM only reaches Q5_K_M (22.13 GB) for the same model. Speed flips that picture hard: the RTX 3090's 936 GB/s is 6.24x the M3 Pro's 150 GB/s, and while this site's calculator credits Apple's unified memory architecture with higher real-world decode efficiency (80% of rated bandwidth versus 65% for a discrete GPU, which narrows the gap somewhat), the RTX 3090 still decodes roughly 5x faster on anything both fit: Llama 3.1 8B at Q4_K_M runs 102.3 tok/s on the RTX 3090 versus 20.2 tok/s on the M3 Pro, and Qwen3 14B at Q4_K_M runs 58.8 versus 11.6 tok/s, both landing close to that same ratio. Even Qwen 3.6 27B, where the M3 Pro gets the better quant, decodes far slower there: 30.8 tok/s on the RTX 3090's Q5_K_M build versus 5.3 tok/s on the M3 Pro's higher-quality Q6_K build. The RTX 3090 is also the only one of the two with a path past 24GB at all: NVLink-paired with a second RTX 3090, it reaches a 48GB pool for 70B-class models that a single M3 Pro, at this 36GB configuration, still can't hold natively.

Bottom line: For raw decode speed, the RTX 3090 isn't close: roughly 5x faster on every model both machines fit, real-time chat-speed territory the M3 Pro doesn't reach at this size. The M3 Pro's case is capacity-per-watt and portability, not speed: it's a laptop chip drawing a fraction of a desktop GPU's power, with slightly more room for a model like Qwen 3.6 27B to run at a meaningfully better quant. Choose the RTX 3090 (or a pair of them) for actual usable local-inference speed, especially past the 20B class; choose the M3 Pro only if portability and power efficiency matter more than tokens per second, and even then, be aware you're trading roughly 5x the decode speed for a modest capacity edge.