Not too long ago, I was lucky enough to snag an elusive Mac Studio off Ebay (an actual legit one). Now, why are these important, and what really sets them apart from any other computer?
Sure, there's a few basic things. You can run AI agents on it 24/7, and the memory pool is big enough that you can build pretty intensive applications with it in parallel at the same time. However, here's the secret sauce behind why I think Apple will win (or at least be a contender) for the local compute arms race.
Fair warning: a bit of science ahead.
The unified memory architecture (which Apple markets as their M-series / Apple Silicon chips) is super powerful. This means the GPU and CPU share memory together, and can sync their operations together better (broadly speaking).
NVIDIA recently released their own unified memory computers (DGX Sparks) which employ this approach - but for most of their history, they had separate consumer graphics cards that would only go up to 8, 12, maybe 24 GB of dedicated video RAM. This card would be paired with an Intel or AMD processor - a different manufacturer.
However, there are a few "gotchas" with Apple's approach in particular.
SCIENCE
The Apple Silicon GPU does not support float64 calculations - only float32. In other words, you get roughly 7-8 decimals with float32, and 16 decimals with float64.
Although the unified memory is incredibly powerful for overall pool access, for scientific computing, there is several orders of magnitude more overhead per tiny operation (CUDA measured in nanoseconds vs PyTorch MPS in much bigger microseconds).
/END SCIENCE
However: because Apple manufactures both the CPU and the GPU for their computers, and has had the unified memory model for years, they have an incumbent advantage here with their large audience. Plainly speaking: for 95% of use cases, if you are using a Mac, you have a categorically advantageous computer.
Peter Steinberger, the creator of OpenClaw, laments the Greek Fire-esque loss of manufacturing high-memory Mac Studios
Before Apple got hit by the RAM shortage, they had 256 GB and 512 GB RAM versions of their Mac Studio available directly on their site for eye-watering cost. They're no longer listed. NVIDIA DGX Sparks are 128 GB apiece, and if you buy a fresh Windows computer off Amazon, I doubt you'll get more than 24 GB of VRAM.
If you want to join this arms race and see what your computer can run locally, this is my favorite site to use: https://www.canirun.ai
Right now, I'm having mine crunch spreadsheets, pull data off my Google Drive and public APIs, read my journals, generate images and short videos for me (courtesy of SenseNova and LTX 2.3), organize my calendar, and automate a dozen or so browsers for me. This is going to be the brain for my genius loci too.
Be well,
Michael Kirsanov