Mac Studio takes on DGX Spark as AMD's MEXT halves server DRAM needs

Mac Studio and DGX Spark go head to head

Tom's Hardware has published a detailed comparison of Apple's Mac Studio and Nvidia's DGX Spark, two machines aimed at people who want to run AI models on their own hardware. According to the publication, the testers put both systems through a range of real-world and synthetic measurements, including large language models on each device.

The comparison matters because both machines carry high price tags, so buyers need to match the hardware to their intended workloads. Tom's Hardware notes that inference engine settings are important when running bigger models, and that Apple's MLX software stack is maturing. The Mac Studio tested uses the M5 Ultra chip, while the DGX Spark is Nvidia's compact AI computer.

This comparison sits within a growing market for local AI machines. Alongside these two systems, buyers can also consider Nvidia's RTX Spark and AMD's Strix Halo platform, both of which target the same enthusiast segment.

AMD's MEXT aims to ease the memory shortage

Memory supply has become a serious problem across the industry as AI datacentre building pushes demand for DRAM chips. Tom's Hardware reports that AMD has shared more detail about MEXT, a predictive memory engine designed to let servers run with half the DRAM they would normally need.

The technology combines a transformer model with the clever use of flash storage. AMD claims the approach keeps roughly 99% of a server's performance, though a separate headline from the publication refers to 95%. Either way, the company says the trade-off in speed is small compared with the memory saved.

AMD is already using the system in production. According to Tom's Hardware, 145 servers currently run with MEXT in place, and the company plans to expand that to 2,000 servers, with the headline pointing to 2027 as the target year.

AI tools that design chips faster

Tom's Hardware also interviewed the Chief Product Officer of Cognichip, a company building AI tools for semiconductor engineering. In September, Cognichip unveiled a set of proprietary models called Artificial Chip Intelligence, or ACI, aimed specifically at chip design work.

The company claims its tools can cut design times from months to days by automating tasks that normally demand heavy manual effort. According to the interview, the tools have not yet appeared in a shipping product, so the claimed gains remain unproven in commercial silicon.

Nvidia GPUs, export controls and Chinese demand

One of the week's most striking stories concerns how Nvidia hardware reaches China despite export controls.

Demand for Nvidia compute in China remains strong, with the publication reporting that more than 6 million units of the export-compliant, GB202-based RTX 5500 Pro have reportedly been ordered by AI labs. Chinese buyers may also need to run enough domestic chips to comply with Beijing's rules.

Some Chinese companies have responded by renting compute in Southeast Asian countries such as Malaysia and Singapore. Tom's Hardware reports that Nvidia has cracked down on these overseas cloud rentals, requiring partners to keep a 51% non-Chinese quota, while black market prices for B300 racks have reportedly reached $2.8 million.

Google expands its TPU line and a new jailbreak rating system

Google has added two chips to its eighth-generation TPU range, according to Tom's Hardware. The TPU 8i Max doubles the specifications of the original TPU 8i by placing two chips in one package, targeting long-context agentic AI work. The TPU 8tx takes a different route, replacing Arm host processors with AMD EPYC chips for specialised security workloads.

On the safety side, the Frontier Model Forum has proposed a Jailbreak Rating Framework. The system would grade how easily an AI model can be pushed past its safeguards to produce harmful content, using four scoring components that cover capability gains, scope and ease of deployment.

The idea resembles the CVSS scoring used for software vulnerabilities. The foundation hopes labs can identify likely jailbreak routes and fix them before they are exploited in the real world.

A week of change at Nvidia

Tom's Hardware frames much of this coverage around one event: Jensen Huang has stepped away from his role as Nvidia's chief executive. The publication notes that the ramifications are wide, given how the company transformed from a gaming GPU maker into an AI business over the past five years.

The stories above, from Chinese GPU demand to competing AI hardware, show how central Nvidia remains to the industry even amid that leadership change. Readers who want the full deep dives will need a Tom's Hardware Premium subscription, which costs $29 per year or from $7 per month.