Apple M5 Ultra Mac Studio outpaces DGX Spark on local LLMs

Tom's Hardware's review of the M5 Ultra Mac Studio, by Andrew E. Freedman, reports the machine's tokens-per-second throughput on local LLMs at almost four times that of Nvidia's DGX Spark across the board, and roughly double the outgoing M4 Max. Prompt-processing speeds on the Apple box also beat DGX Spark.

The reviewer points at memory bandwidth. The M5 Ultra pushes 1.2 TB/s of unified memory against the DGX Spark's 273 GB/s, and generation speed on local models is bandwidth-bound. On the CPU side, the sample scored 3,770 single-core and 52,293 multi-core on Geekbench 7, edging an AMD Ryzen Threadripper 9980X in video transcoding tests by eight seconds. A higher-end Threadripper Pro 9985WX still scored 56,118 multi-core.

"Combined with the blazing-fast single- and multi-threaded performance of the 36-core CPU, the Mac Studio with M5 Ultra looks to be ready for anything one might want to do with local or agentic AI workloads," Freedman writes.

The top configuration tested carries a 36-core CPU, 80-core GPU, 256GB of RAM and a 4TB SSD at $12,299. The base Mac Studio M5 Ultra starts at $5,499 with a 30-core CPU, 64-core GPU, 96GB of memory and a 1TB SSD; moving to 256GB adds $4,000. Apple has not published a price for the 512GB unified-memory option, which is slated for late October.

DGX Spark's local-inference pitch has been a running thread through our Nvidia coverage this quarter, next to a broader run of inference stories making the bandwidth-versus-cost math suddenly load-bearing.