The 1.2 TB/s Mac Studio asks you to pick your model size before you buy it

Divide the bandwidth by the capacity. 1.2 TB/s over 256GB is about 4.7 full reads of memory per second ^[20]. If decode reads every weight once per generated token, as a dense model does, then a model sized to fill that memory tops out near 4.7 tokens per second. Attention over the context, sampling and framework overhead take their cut from there. Use half the memory and the ceiling roughly doubles.

Apple puts UltraFusion die-to-die bandwidth above 4.4 TB/s ^[3], roughly 3.7 times the external memory figure ^[19]. On a quad-die part assembled from two dual-die M5 Max SOCs ^[2], that ratio says DRAM is the limit during generation, not the seam between dies. In a memory-bound decode, the rest of the silicon waits: the 80 GPU cores each carry a neural accelerator, and there is a 32-core neural engine alongside them ^[10].

The review's headline says the machine "outpaces DGX Spark and Threadripper" ^[9], and its verdict lists "1.2 TB/s memory throughput is excellent for local AI models" as a pro ^[8]. The published figures are a Geekbench 7 win ^[12] and Cinebench clock estimates of 4.6 GHz single-core and 4.3 GHz multi-core ^[11]. Tom's Hardware says it compared the system against workstations instead of desktops because it does not test many workstations ^[17]. For a local-serving comparison to move to your stack, the other box has to run the same quantization, the same context length and the same runtime. A bandwidth ratio does not survive a switch from a 4-bit dense model to a mixture-of-experts with few active parameters. The review does not price the machine against rented GPU hours, so the utilisation maths stays with the buyer.

Sustained inference in this enclosure is a thermal question. The chassis is 7.7 by 7.7 inches and 3.7 inches tall ^[24]. The M5 Ultra version weighs 8 pounds, two more than the M5 Max variant, because Apple swaps the thin aluminium stack and copper heat pipe for a copper fin stack and a copper vapour chamber ^[15]. Two pounds of copper is the cost of holding those clocks in a box that size.

Apple discontinued the Mac Pro earlier this year, leaving the Mac Studio as its top-end desktop ^[13], and this is the only machine Apple currently puts the M5 Ultra inside ^[25]. There is a Kensington lock slot, on the bottom, which needs a special adapter to use ^[16].