Buying one of these is an unbelievably bad deal
You could literally get two 256GB M5 Ultra Mac Studios for less money and connect them over thunderbolt 5 RDMA and get significantly faster AI compute, not to mention twice the storage
0nly kinda makes sense if your targeting a 2 TB+ cluster for $85k~
I imagine this decision was made to exclusively target enterprise customers, but once you're spending near $100k on local AI, I'd imagine most businesses willing to drop this much money would would prioritize the much higher AI compute that's Nvidia alternatives would offer over the energy efficiency of Macs
Even if technically $85k~ is cheap compared to standard Nvidia or AMD alternatives if all you want is capacity, models that large are not going to be fast even in an RDMA cluster of these things
So it only makes sense if you're planning on buying four or more of these things and are okay with frontier models running very slowly, probably 10-20 tokens per second on frontier models that can actually take advantage of the memory
However this is still mostly a bad deal since projects like Strata are becoming more popular, which allow these large MoE models to intelligently stream experts from fast PCIe 5.0 storage into system memory and VRAM and keep them in faster memory based on how frequently they are used
If all you care about is capacity, advancements in software are making if possible to run extremely large models for much cheaper using 2+ TB 14 GB/s SSDs combined with 256-512gb of system memory and more modest amounts of vram on GPUs connected over fast PCIe 5.0 lanes
A system like this would likely cost somewhere around $20k and produce similar speeds to the 512 gb M5 ultra cluster for processing one request at a time
Parallel requests might take a harder hit than the M5, but not as much as you might expect, and at a quarter of the price, you could literally get more overall parallel throughput by building multiple of these hybrid systems and still come in cheaper than if you were to have bought the M5 ultra cluster
The only remaining actual strong suit of the M5 Ultra cluster would be fine tuning large models, which would be really poorly suited for the hybrid system, but would be possible yet slow on the Mac cluster
Fine tuning 1-2 TB models is genuinely the only hyperspecific reason that Apple thinks they can get away with pricing the 512 GB option this ludicrously expensive compared to the 256 GB option
Anyway sorry for geeking out, I promise my next tweet will be something racist 🙏