Mac Studio M5 Ultra Beats DGX Spark on Speed and Energy in Local AI Tests

A single $12,300 Mac Studio with Apple's M5 Ultra finishes the AgentPerf-Local agentic benchmark in just under half the time of an Nvidia DGX Spark-class machine, and it consumes about 16% less task energy doing it. That is the headline result of extended local AI testing by Tom's Hardware, which pits Apple's flagship unified-memory workstation against GB10-based systems now priced above $10,000 each.

The tested Mac Studio pairs a 36-core CPU and 80-core GPU with Neural Accelerators for matrix math. It carries 256GB of unified memory in the reviewed configuration, expandable to 512GB, with 1.2 TB/s of memory bandwidth. Nvidia's GB10 offers 48 Blackwell SMs and 273 GB/s of bandwidth — roughly one quarter of the Mac's decode bandwidth. Dell's Pro Precision GB10 with a 4TB SSD now sells for around $10,000, making the head-to-head comparison legitimate on price.

What did the agentic tests show?

The tests used Artificial Analysis' AgentPerf-Local benchmark, which replays recorded inference traces across agentic workloads. Most of its work sits in prompt processing rather than token generation, favoring systems with heavy raw compute. Both platforms ran the dense Qwen 3.8 27B model through llama.cpp, with native multi-token prediction enabled.

The Mac Studio finished the fixed workload in under half the Spark's time. It also consumed roughly 16% less task energy measured in kilojoules at the wall. The Mac's energy-delay product came in at under half the DGX Spark's figure, combining speed and efficiency into a single metric.

How do the huge models run?

Qwen 3.8 27B does not exercise 256GB of memory. The testers turned to two larger open models near the top of capability benchmarks: DeepSeek V4.1 Flash 1017, run as a two-bit Unsloth quantization, and GLM 5.3 Flash in four-bit form. Neither fits on a single 128GB DGX Spark. That constraint forces clustering.

Two Dell GB10 systems with 4TB SSDs cost $15,000 to $20,000 or more, roughly 25% above the Mac Studio even at the lower bound. Lenovo moved its ThinkStation PGX systems to "coming soon" status during testing, and analysts expect GB10 pricing to stay elevated as the DRAM and NAND supply crunch deepens.

The results split by model. For DeepSeek V4.1 Flash 1017, the Mac Studio must use llama.cpp, and it cannot hold both the model weights and the DSpark speculative decoding model even at two-bit quantization. The Spark cluster runs a community serving recipe from Mia's AI Lab with the TensorFold engine and speculative decoding. The Mac needs 10 to 25 minutes to digest the longest contexts tested; the cluster finishes in under five minutes. A pair of Sparks is the clear winner for that model today.

Does the software gap close?

GLM 5.3 Flash narrows the gap. Its multi-token prediction needs no separate drafting model, so the Mac gains a decode boost without memory pressure. The Spark cluster still delivers faster prompt processing across the board, while the M5 Ultra's bandwidth advantage yields higher decode tokens per second.

A TensorFold serving recipe exists for GLM 5.3 Flash on Apple Silicon, and it transforms the result. Decode throughput runs three times faster on average than llama.cpp, and prompt processing approaches the cluster's numbers. The testers expect improved recipes for DeepSeek on Apple Silicon to arrive as community developers target the hardware, because the performance headroom clearly exists.

What should buyers do?

Lead times for the Mac Studio stretch 13 to 15 weeks, into February 2027. The reviewers' verdict is split. Buyers who need one well-rounded workstation with a fast CPU and GPU in a quiet chassis should order the Mac and wait for MLX software to mature. Buyers spending more than $10,000 purely for local inference get the memory capacity, compute and broad model compatibility they need from a two-Spark cluster, and that software assurance is worth the lower raw performance ceiling.

The broader dynamics point one direction: with DRAM and NAND supply tightening and GB10 prices rising, the Mac Studio stays appealing only as long as Apple holds its own prices flat.