Editor's Choice, With Caveats: Tom's Hardware's M5 Ultra Mac Studio Review Beats DGX Spark at Local LLMs

The reviews are in, and Apple’s bet on desktop AI hardware just got its first independent seal of approval. On September 21, 2026, Tom’s Hardware published its full review of the Mac Studio with M5 Ultra — awarding it an Editor’s Choice — and the benchmark numbers give Nvidia’s much-hyped DGX Spark a bloody nose in the one market both machines are chasing: running large language models locally.

The headline finding is blunt. In Tom’s Hardware’s local AI testing, the M5 Ultra’s prompt processing speeds are faster than Nvidia’s DGX Spark’s, and its tokens-per-second throughput is double that of the M4 Max and almost four times higher than the DGX Spark across the board. The test model was Qwen 3.8-27B-Q4_K_M — a dense model chosen deliberately, because dense inference is memory-bandwidth-bound, and bandwidth is exactly where Apple’s silicon philosophy pays off.

Why bandwidth wins

The M5 Ultra is Apple’s first quad-die chip: two dual-die M5 Max packages fused over the UltraFusion interconnect, which Apple says pushes over 4.4 TB/s of die-to-die bandwidth. The configuration matters more than the core counts — the review unit paired a 36-core CPU and 80-core GPU with 256GB of LPDDR5 unified memory on a 1.2 TB/s bus, double the memory bandwidth of the M4 Max that Tom’s Hardware previously tested for AI workloads. Every one of the 80 GPU cores carries its own neural accelerator, and a 32-core Neural Engine sits alongside for good measure.

That 1.2 TB/s figure is the whole story for local LLMs. Token generation speed on a dense model is largely a function of how fast you can stream weights through memory. The DGX Spark’s GB10 package offers 128GB at roughly 273 GB/s; the M5 Ultra offers up to 256GB (512GB arriving in late October) at more than four times the bandwidth. Nvidia’s box wins on raw compute and CUDA’s software moat, but for a single practitioner loading one very large model, Apple now has the more coherent machine.

The numbers beyond AI

The review frames the Studio as the machine inheriting the Mac Pro’s mantle after that product’s discontinuation earlier this year. On general productivity it mostly justifies the crown:

• Geekbench 7: 3,770 single-core / 52,293 multi-core — the clear winner against the workstation comparison set, despite fewer cores than its x86 rivals.
• Handbrake: a 4K-to-1080p transcode in 1:04, eight seconds ahead of the 64-core AMD Ryzen Threadripper 9980X.
• File copy: 2,983 MB/s moving 25GB of files, beating both the M3 Ultra and M4 Max from the previous generation.
• Sustained load: across a 10-run Cinebench 2026 stress test, scores ran 17,825–18,069 with the CPU averaging 78.29°C — remarkably consistent thermals for a 7.7-inch-squared, 8-pound chassis sipping from a 480W maximum power budget.

It isn’t universally dominant. CPU-based Blender rendering falls behind both current and previous Threadrippers and the 56-core Xeon w-3495X, and gaming remains the Studio’s weakest suit — 66 FPS in Cyberpunk 2077 at 1080p with ray tracing, roughly RTX 5070-class, and unplayable at 4K while RTX 5090 systems held 59 FPS. The verdict’s cons list is honest: pricey memory and SSD upgrades, and zero post-purchase upgradeability, since the RAM is soldered.

The economics of a $12,299 desktop

As tested — top-bin M5 Ultra, 256GB, 4TB SSD — the review unit costs $12,299, and Tom’s Hardware is candid that this makes it “effectively an enterprise computer.” The pricing ladder below it is steep but real: $2,499 for a base M5 Max with 36GB, or $5,499 for the cheapest M5 Ultra with 96GB. The $4,000 jump from 96GB to 256GB of RAM and the $1,500 4TB SSD are where the bill balloons.

Two data points from the review put that price in context. First, a 64-core Threadripper 9980X alone — a chip that beats the M5 Ultra in Blender — costs around $5,000 before you add memory, storage, or a case. Second, demand is doing the talking: as of publication, the 256GB configuration is backordered 16–18 weeks, and third-party rack mounts for the unchanged Studio chassis already exist for homelab and server-farm buyers. The reviewer’s verdict for the local AI crowd is unusually direct: “If you are interested in local AI, this will definitely do the job.”

The bigger picture

This review lands as a checkpoint in a broader shift. Mac Studio supply was already strained this year by an “OpenClaw-fueled ordering frenzy” that pushed delivery of high-unified-memory units out six days to six weeks, and Apple has since pulled the $4,000 512GB upgrade option from the store while supply catches up. A whole cottage industry — from MLX cluster tooling to performance trackers maintained in Reddit threads — has grown around treating Studios as disposable inference nodes.

The Tom’s Hardware data sharpens the argument that was previously built on Apple’s own marketing claims (the company had promised up to 4.3x the peak AI compute of the M3 Ultra and up to 9.8x faster prompt processing versus M1 Ultra). Independent numbers now confirm the shape of the story: for single-box, large-model, bandwidth-hungry local inference, Apple silicon is the strongest consumer-priced option on the market — and by a wider margin than many expected.

The caveats are structural, not incidental. No CUDA means the entire NVIDIA ecosystem of kernels, tooling, and enterprise software remains out of reach. Soldered memory means capacity is a purchase-time decision, not an ops decision. And a 512GB ceiling in late October still trails what a multi-GPU x86 box can hold, even if the price per usable gigabyte gets ugly fast on both sides.

But for the practitioner who wants a frontier-adjacent open-weights model — a 27B dense model screaming along at 4x DGX Spark throughput, or something far larger loaded into 256GB of quiet unified memory — sitting on a desk rather than behind an API endpoint, the review’s conclusion is hard to argue with. The Mac Studio with M5 Ultra is the strongest single-box answer to that question anyone has shipped yet, and it took an independent lab, not a keynote, to prove it.

[Sources are rendered from the frontmatter above. Tom’s Hardware review by Andrew E. Freedman, published September 21, 2026.]