Which Mac for local LLM research?
Bear with me on this one. I currently own (personally) a 16" M5 Max 128gb/4tb machine. It's my daily driver. I run local models on this fairly regularly, and it works great -- but does get very warm when doing long duration inference.
I recently started a new job as an AI Architect and I've convinced my VP that I should have a company owned machine with enough power to do R&D using local LLMs. I can pick either another 16" M5 Max 128gb/2Tb machine, or I can instead go with an M3 Ultra Studio 32/80 96gb/2Tb for the same price.
I know the cooling of the Studio is going to be hugely beneficial, and memory bandwidth is slightly better (as is the GPU count).. and I can continue using my *personal* Macbook to ssh/screen share into the Studio, making it seem, on paper at least, that the Studio is a no-brainer.
My concern is memory. 96gb is likely fine for most of the upcoming LLM models. It seems either a model requires a 512gb machine, or it can run on something more pedestrian (ie 32/64gb) with little overlap.
I need to act while I have this opportunity -- I don't know if the budget will exist for an M5 Ultra if/when one is released. Even IF an M5 Ultra is released with capacity for 128gb/256gb, it'll likely be too expensive for my department budget anyway.
Thoughts?
Update - decided to request the 96gb M3U 80 core Studio. Thanks everyone for your input!