M5 Ultra vs RTX 5090: 256GB unified memory takes on 32GB VRAM for local LLMs
Hugging Face · youtube · 2026-10-15
With Apple's M5 Ultra Mac Studio now shipping, Hugging Face compares it against NVIDIA's RTX 5090 for running LLMs locally.
- Specs: the RTX 5090 has 32 GB VRAM while the M5 Ultra offers up to 256 GB of unified memory, enabling far larger models to fit.
- Benchmarks: head-to-head tests on Qwen3.8-27B covering prompt processing and token generation, plus multi-token prediction (MTP) speedups.
- Key formula: tokens per second ≈ bandwidth / model size — memory bandwidth drives generation speed.
- References: MacStories' M5 Ultra review ran identical prompts on both machines, and llama.cpp Q4KM results on the RTX 5090 are public on the HF discussion thread.
More from Infra
- Neocloud stocks fall 12% in a day after CoreWeave's bond selloff — SumitGup · 2026-10-15
- TRELLIS.2 image-to-3D runs at full 1536 resolution on a 16 GB card via tiling patches — over-yonder-tech · 2026-10-15
- ServeTheHome tours AWS's secretive networking lab: timing, optics and hyperscale — firstadopter · 2026-10-15
- vLLM adaptive verification now works head-free, using draft logits with online logistic calibration — vllm_project · 2026-10-15
- One R9700 running Qwen3.6-35B did $426.80 of equivalent Opus coding work in a week — supracode · 2026-10-15
- Eric Jang's three signals of AI overinvestment: compute repricing, open-source catch-up, robot cost halving — ericjang11 · 2026-10-15