Side by side
| M5 Ultra | RTX PRO 6000 | |
|---|---|---|
| What it is | Mac Studio (M5 Max or M5 Ultra) | A graphics card for a desktop PC |
| Memory sizes | 96, 256, 512 GB | 96 GB |
| Memory bandwidth | 1,200 GB/s | 1,792 GB/s |
| For models, at most | about 384 GB | about 95 GB |
| Models that run well | 61 of 73 | 50 of 73 |
| Measured: writing | 179 tok/s | 274 tok/s |
| Measured: reading a prompt | 4,945 tok/s | 14,855 tok/s |
Measured rows: each machine’s fastest result in llama.cpp’s public benchmarks with Llama 2 7B Q4_0, writing 128 tokens (tg128) and reading a 512-token prompt (pp512), in tokens a second. The Mac ran Metal, the RTX PRO 6000 CUDA. Apple Silicon results · RTX PRO 6000 results For models we count the RTX PRO 6000’s own 96 GB less about 1 GB, and about 75% of the Mac’s unified memory, the macOS default.
Where each one wins
- Writing speed. The RTX PRO 6000 reads memory about 49% faster than the M5 Ultra (1,792 vs 1,200 GB/s), so it writes about 49% faster with any model that fits both.
- Model size. The M5 Ultra has about 384 GB for models, the RTX PRO 6000 about 95 GB. 61 of 73 popular open models run well on the M5 Ultra, 50 on the RTX PRO 6000.
- Reading long prompts. In llama.cpp’s measured test the RTX PRO 6000 reads a prompt about 3× as fast (14,855 vs 4,945 tokens a second). Reading a prompt depends on GPU compute rather than memory bandwidth, so this sets the wait before the first word with a long document or codebase.
- Software. The RTX PRO 6000 runs NVIDIA’s CUDA, which most AI software supports first, including training code and many image and video tools. On a Mac you run models with MLX, llama.cpp’s Metal backend or PyTorch’s MPS backend instead.
- Around it. A graphics card needs a PC around it, and its own 96 GB is the limit: a model that doesn’t fit has to run partly from the PC’s much slower system memory, or across a second card.
- The Mac. It’s also an everyday computer, and it runs MLX Harness and MLX Node, which are made for Apple Silicon.
Speed with popular models
Tokens a second for one conversation, each chip at its largest memory size (512 GB and 96 GB), with the best version of the model that fits.
| Model | M5 Ultra | RTX PRO 6000 |
|---|---|---|
| Llama 3.1 8B Instruct8B | 64–113 8-bit | 95–169 8-bit |
| gpt-oss-20b21B · 3.6B active | 284–505 native 4.25-bit | 424–755 native 4.25-bit |
| Qwen3.6 35B A3B35B · 3B active | 169–300 8-bit | 252–448 8-bit |
| Qwen3.8 27B27B | 19–33 8-bit | 28–50 8-bit |
| Llama 3.3 70B Instruct70B | 14–24 4-bit | 11–19 8-bit |
| gpt-oss-120b117B · 5.1B active | 200–356 native 4.25-bit | 299–531 native 4.25-bit |
| Qwen3.5 122B A10B122B · 10B active | 51–91 8-bit | 144–256 4-bit |
| DeepSeek V4 Flash 0731284B · 13B active | 39–70 8-bit | too big |
Questions
- Is the RTX PRO 6000 faster than the M5 Ultra for LLMs?
- The RTX PRO 6000 reads memory about 49% faster than the M5 Ultra (1,792 vs 1,200 GB/s), so it writes about 49% faster with any model that fits both. In llama.cpp’s measured test the RTX PRO 6000 reads a prompt about 3× as fast (14,855 vs 4,945 tokens a second). Reading a prompt depends on GPU compute rather than memory bandwidth, so this sets the wait before the first word with a long document or codebase.
- Which can run bigger models, the M5 Ultra or the RTX PRO 6000?
- The M5 Ultra has about 384 GB for models, the RTX PRO 6000 about 95 GB. 61 of 73 popular open models run well on the M5 Ultra, 50 on the RTX PRO 6000.
- Can MLX Node run on the RTX PRO 6000?
- No. MLX Node and MLX Harness run on Apple Silicon Macs only, with Apple’s MLX framework.
More comparisons
The M5 Ultra for local AI · Compare every model · The best Mac for local LLMs
How it’s worked out
Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.
RTX PRO 6000 figures: NVIDIA: RTX PRO 6000 Blackwell Workstation Edition, NVIDIA: RTX PRO Desktop GPUs (compare table).