What the benchmarks actually show
Tom’s Hardware tested a 256GB Mac Studio against Nvidia’s DGX Spark (and an M4 Max Mac Studio). Its local-AI testing used a single model, Qwen 3.8-27B at Q4_K_M, run in llama.cpp with multi-token prediction off. The Mac generated 46.27 tokens per second against 11.8 for the DGX Spark at zero context, roughly four times faster, and its prompt processing was faster than the Spark’s as well.
These are the publication’s measurements, not independently replicated results. Because the review tested one model under one engine configuration, it does not establish how other models or inference engines would rank.
Memory capacity is not a performance guarantee
Apple’s specifications establish the Mac’s considerable hardware headroom: M5 Ultra offers up to a 36-core CPU, an 80-core GPU, 512GB of unified memory and 1.2TB/s memory bandwidth. Those maximum specifications should not be confused with the memory capacity of the tested configuration.
Nvidia lists DGX Spark with a 20-core Grace CPU and up to 128GB of unified memory. Its platform documentation explicitly describes two 128GB systems providing 256GB of aggregate capacity, with ConnectX networking supporting multi-system operation. That establishes the capacity comparison—not equivalence to one machine’s local memory performance.
Clustering also deserves a qualification: it is not exclusive to Nvidia. Apple documents Thunderbolt 5 and RDMA support for clustering Mac Studio systems. However, Apple’s vendor-reported scaling claims are not measurements of other models or configurations, so they cannot settle those comparisons by proxy.
The takeaway: distinguish three questions—whether a model fits, whether an engine supports it, and how quickly the complete workload finishes.
How to evaluate a local AI workstation
For readers planning their own comparison, llama.cpp’s official benchmarking documentation provides a useful starting point: llama-bench distinguishes prompt processing, text generation and combined prompt-plus-generation tests. It also records configuration details such as the model, backend, batching parameters and build information. A lone tokens-per-second number leaves too much unsaid.
A sensible evaluation plan is therefore to:
- Specify the exact model and quantization before comparing machines.
- Measure input processing and output generation separately, then test the combined workload.
- Record the software build and settings, rather than treating the engine’s name as a complete configuration.
- Define success around your task: an interactive assistant and a long-document processing job need not reward the same performance characteristics.
These are proposed evaluation criteria, not additional benchmark findings. They follow the distinctions exposed by the benchmark tooling.
What this means for Windows users
There is also a practical distinction between replacing a desktop and adding an AI compute appliance. Nvidia’s DGX Spark documentation describes NVIDIA Sync as a remote-access utility for Windows, macOS and Ubuntu. Its product page says users can launch tools including VS Code, Cursor and AI Workbench remotely. A Windows workstation can therefore remain the working environment while Spark supplies separate compute resources.
That does not make Spark the automatic winner. It makes workflow compatibility another item on the scorecard.
The purchasing question should start with “Can this exact software stack run my workload well?”, not “Which specification looks most impressive?” Benchmark the recipe you intend to use—not the performance you hope a future update might deliver.
Correction (November 10, 2026): An earlier version of this article wrongly stated that Tom’s Hardware tested the Mac Studio against Dell Pro Precision GB10 hardware, and attributed AgentPerf-Local, DeepSeek V4.1 Flash 1017, and GLM-5.3-Flash findings to Tom’s Hardware. Tom’s Hardware’s September 21, 2026 review tested the 256GB Mac Studio against Nvidia’s DGX Spark (and an M4 Max Mac Studio), ran Qwen 3.8-27B in llama.cpp with multi-token prediction off, and did not test Dell hardware, DeepSeek, GLM-5.3-Flash, or TensorFold. The DeepSeek and GLM items have been removed and the Qwen finding corrected to the review’s published figures.