Nvidia's DGX Spark and Apple's Mac Studio are the subjects of a head-to-head comparison aimed at buyers who want PCs for running local AI models. Tom's Hardware ran real-world and synthetic benchmarks on both machines and also tested large models on each. The publication noted that prices for both devices are at eye-watering levels, making it more important than ever for buyers to choose correctly for their intended workloads. The comparison sits alongside other options on the market, including the recently released RTX Spark and AMD's Strix Halo platform.
AMD has disclosed more detail about MEXT, a predictive memory engine the company says can deliver 99% of a server's performance while using half the DRAM. The system is powered by a transformer model and makes clever use of flash storage, according to the report. AMD is already running 145 servers with MEXT in place and aims to expand that figure to 2,000 next year. The disclosure comes as DRAM capacity constraints remain an industry-wide sticking point, with AI buildout accelerating demand for dies and costs spiraling across client and enterprise hardware.
Tom's Hardware also interviewed Cognichip's Chief Product Officer about how the company is rolling out AI-powered chip design tools. In September, Cognichip revealed a suite of proprietary AI models called Artificial Chip Intelligence, or ACI, developed specifically for semiconductor engineering. The tools have not yet been used in a shipping product, but they could substantially improve chip development productivity by automating labor-intensive tasks.
The Jensen Huang scandal deepened, with the now-former Nvidia CEO embroiled in a series of alleged extramarital affairs and having since resigned from Nvidia. Much of the past week's coverage centered on his relationship with alleged Chinese Spy Alice Huang, and the report also explored his unlikely relationship with Kang Wei, Ms. Huang's husband, who recently landed in hot water with China's Securities Regulatory Commission.
Nvidia-based compute remains difficult to find in China, especially as AI labs line up to order racks and newer cards like the RTX 5500 Pro. More than 6 million units of the export-compliant GB202-based AI accelerator have reportedly been ordered. Chinese AI labs chasing Nvidia's tech may still have to pay a premium and ensure they run enough homegrown silicon to avoid falling afoul of Beijing's rules, and some companies have moved to renting compute in Southeast Asian countries like Malaysia and Singapore to get around export controls.
Google's 8th-gen TPU lineup expanded with two new chips designed for long-context agentic AI and security. The two chips, TPU 8i Max and TPU 8tx, signal further diversification across Google's eighth-generation TPU stack. TPU 8i Max doubles the specs of the original TPU 8i, while TPU 8tx makes a more interesting change to AMD's EPYC to serve its use in specialized security workloads.
The Frontier Model Foundation has proposed a Jailbreak Rating Framework that aims to classify the type and severity of a possible LLM jailbreak. The framework has four scoring components, and the foundation aims to identify possible jailbreak vectors and inform labs of them before they happen in the real world, where the consequences could be disastrous. The proposal addresses how easy it might be for an AI model to break through its safeguards and generate potentially harmful content.
More AI news from TechManNews.
