2x RTX 3090 vs Mac M3 Ultra 256GB for local LLMs
Pick the Mac M3 Ultra 256GB if you want to run bigger models (it runs 49 of the top 60 vs 43); pick the 2x RTX 3090 if the models you use already fit in it and you want more tokens per second.
- The Mac M3 Ultra 256GB runs more of the 60 most-downloaded models fully in memory at 4-bit or better: 49 vs 43, because it has more usable memory (about 192 GB vs 46 GB).
- At near-lossless 8-bit or better, the Mac M3 Ultra 256GB runs 45 of them vs 43.
- At the same quant, the 2x RTX 3090 has a 1.1× higher speed ceiling (memory bandwidth 936 vs 819 GB/s). The speed table below shows each card's best quant that fits, so a card with more memory can show a slower, higher-quality setting.
Specs that matter for LLMs
| 2x RTX 3090 | Mac M3 Ultra 256GB | |
|---|---|---|
| Memory (VRAM) | 2×24 GB | 256 GB unified |
| Usable for a model (our estimate) | 45.6 GB | 192.0 GB |
| Memory bandwidth | 936 GB/s | 819 GB/s |
| Launch MSRP | $2,998 | — |
| Launch $ per GB of memory | $62 | — |
| Top 60 models that fit (4-bit or better, 8K ctx) | 43 | 49 |
| …of those at 8-bit or better | 43 | 45 |
| Largest top-60 model that fits | Qwen3.5 35B A3B | Hy3 |
| Offload headroom: more top-60 models via llama.cpp CPU offload (≥3 tok/s est., up to 128 GB RAM) | not modelled for multi-GPU setups | n/a: one shared memory pool |
Top 60 = the 60 most-downloaded open models we track. Launch MSRP from vendor announcements; Macs and workstation parts have no single launch price, so none is shown. How we calculate.
Estimated speed on popular models
Best quant that fits at 8K context, with a tokens/sec ceiling from memory bandwidth (real speed is typically 50–80% of it). "Offload" means part of the model sits in system RAM. Estimates, not benchmarks.
| Model | Needs at 4-bit | 2x RTX 3090 | Mac M3 Ultra 256GB |
|---|---|---|---|
| Qwen3 8B | 6.5 GB | F16 · ~57 tok/s | F16 · ~50 tok/s |
| Llama 3.1 8B Instruct | 6.2 GB | F16 · ~58 tok/s | F16 · ~51 tok/s |
| Qwen3 14B | 10.3 GB | F16 · ~32 tok/s | F16 · ~28 tok/s |
| gpt-oss 20B | 11.8 GB | F16 · ~338 tok/s | F16 · ~296 tok/s |
| Gemma 3 27B IT | 17.3 GB | Q8_0 · ~33 tok/s | F16 · ~15 tok/s |
| Qwen3 32B | 21.1 GB | Q8_0 · ~27 tok/s | F16 · ~12 tok/s |
| Qwen3 Coder 30B A3B Instruct | 18.8 GB | Q8_0 · ~293 tok/s | F16 · ~137 tok/s |
| Llama 3.3 70B Instruct | 42.7 GB | Q3_K_M only (below 4-bit) | F16 · ~6 tok/s |
| gpt-oss 120B | 59.6 GB | No | F16 · ~255 tok/s |
Models only one of them runs
Only the 2x RTX 3090
None.
Only the Mac M3 Ultra 256GB
- Qwen3.8 Flash Next (180B, 112 GB at 4-bit)
- gpt-oss 120B (117B, 59.6 GB at 4-bit)
- DeepSeek V4 Flash (291B, 167 GB at 4-bit)
- GLM 4.5 Air (110B, 70.6 GB at 4-bit)
- MiniMax M2.5 (229B, 132 GB at 4-bit)
- Hy3 (299B, 173 GB at 4-bit)
More comparisons
See every model on the 2x RTX 3090 or the Mac M3 Ultra 256GB, or try your own settings in the VRAM calculator.