Mac vs PC for local AI
PC with NVIDIA = more speed per euro on models that fit in VRAM. Mac (or Strix Halo / DGX Spark) = huge unified memory, so very large models fit at all, just slower.
| Model | Mac (tok/s) | PC (tok/s) |
|---|---|---|
| gpt-oss 20B | 115.29 (Mac Studio M3 Ultra 512GB (Geerling test)) | 282.51 (Single RTX 5090 32GB) |
| gpt-oss 20B | 92.36 (Apple M4 Max 36GB) | 161.77 (Single RTX 3090 24GB) |
| gpt-oss 120B | 79.68 (Mac Studio M2 Ultra 76-core GPU, 192 GB) | 196.31 (RTX Pro 6000 96GB workstation card) |
| Llama 3.1 70B | 14.08 (Mac Studio M3 Ultra 512GB (Geerling test)) | 3.1 (Ryzen 9 7900X + RTX 4090) |
| DeepSeek R1 671B | 19.89 (Mac Studio M3 Ultra 512GB (Geerling test)) | 8.97 (EPYC + single RTX 3090 (hybrid)) |
Every number is a measured generation speed from the linked build page and its source. Engines and quants differ between sources. The Llama 3.1 70B row shows what happens when a 70B model does not fit a 24 GB card (3.1 tok/s with offload) vs a 512 GB Mac (14.08). The DeepSeek R1 PC row is a single 3090 + EPYC with most of the model in system RAM at Q2_K.
Rule of thumb
- Model fits in 24–32 GB (Qwen3.6 35B-A3B, Gemma 4, gpt-oss 20B)? An RTX 3090/4090/5090 PC is faster.
- Want 120B+ (gpt-oss 120B, Qwen3.5 122B, Nemotron 3 Super) without multi-GPU? 128 GB+ unified memory is the simple route.
- Check exact fit with the VRAM calculator.
Not enough VRAM? Rent before you buy. Run a 120B model on a rented 24–96 GB cloud GPU by the hour and see the real speed before spending €1,000+. Try it on Runpod
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.
Referral link: if you sign up with Google and add $10, Runpod gives you and RigBase bonus credit (terms). It never affects our data.