Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Spark
But if you're thinking about dropping over $10K for the purpose of building a local AI inference setup, is a Mac Studio right for you?
Apple's latest Mac Studio, equipped with the M5 Ultra SoC, has demonstrated its prowess in the realm of artificial intelligence, boasting a 36-core CPU and an 80-core GPU with Neural Accelerators for matrix math operations commonly used in AI compute. Highly sought after, this powerful machine offers up to 512GB of unified memory, which, coupled with its impressive specifications, has many AI enthusiasts eagerly awaiting its shipment, a process that takes 13 to 15 weeks.
But is this system worth its hefty price tag, potentially upwards of $12,300 for some users? To answer this question, we conducted a series of tests utilizing cutting-edge large models like DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash. Our initial review focused on the Mac Studio's performance with the popular Qwen 3.8 27B, but we quickly realized that the 256GB of memory proved insufficient to fully exercise these larger models.
To thoroughly test the Mac Studio's capabilities, we incorporated new benchmarks and compared its performance against Nvidia's DGX Spark. A key benchmark we employed was the AgentPerf-Local benchmark, which assesses agentic AI performance and power efficiency by replaying recorded inference traces across a range of AI workloads.
The test primarily focuses on prompt processing rather than token generation, favoring systems with plentiful raw compute capabilities, a trait both the Mac Studio and the DGX Spark possess. The Mac Studio's 1.2 TB/s of memory bandwidth provides it with over four times the decode bandwidth compared to the DGX Spark. However, raw specs do not always equate to superior performance; real-world results often diverge from paper specifications.
To further evaluate the Mac Studio's efficiency, we utilized llama.cpp as our inference backend for both systems, enabling Qwen 3.8 27B's native multi-token prediction (MTP) support for a speed boost in decode operations. The results were compelling: the Mac Studio completed the AgentPerf-Local test in less than half the time of the DGX Spark, consuming approximately 16% less task energy.
To put this into perspective, a single DGX Spark-like system, such as Dell Pro Precision GB10, costs around $10K, while the Mac Studio is priced at $12K. When comparing the systems' performance, energy consumption, and cost, the Mac Studio emerges as a formidable contender, delivering impressive results at a more accessible price point.
Written by urgent.news from Tom's Hardware's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
