Mac Studio M4 Ultra vs RTX 4090 for Local LLMs
Two very different answers to the same question. Memory size versus speed.
Mac Studio M4 Ultra 256 GB exposes ~192 GB unified memory to the GPU at 1092 GB/s. RTX 4090 has 24 GB at 1008 GB/s.
For 7B-13B models, the 4090 is dramatically faster, 2-3x on tokens per second. For models that do not fit in 24 GB (70B, 405B, DeepSeek V3), the Mac is the only single-machine option.
Power: Mac pulls 30-100W during inference, the 4090 alone pulls 400+. Noise: Mac is silent, a 4090 in a tower is loud.
If your workload is one user, small models, fastest possible: 4090. If your workload is fitting models that NVIDIA consumer cards cannot fit, on a quiet desk machine: Mac. Different tools, different jobs.