Reviews from Notebookcheck and PC Gamer found that external GPU enclosures connected via Thunderbolt suffer meaningful performance losses compared to desktop PCIe slots, ranging from 7 to 32 percent depending on resolution and card. PC Gamer found an RTX 4070 Ti ran about 25 percent slower over Thunderbolt than OCuLink, with worse stuttering and lower frame-rate consistency, and found that pairing an eGPU with a faster RTX 4090 yielded diminishing returns versus a full desktop connection.
Reuters reported that SK Hynix is in early talks with Intel to produce memory chips in the United States, potentially through leasing space at Intel's Ohio plant or forming a joint venture involving cloud companies. Shares of both firms rose in premarket trading following the report, though sources cautioned no final decisions have been made.
USB-C hubs can substitute for docking stations by adding extra ports to a laptop, but they draw power and bandwidth directly from the host computer rather than a dedicated outlet or bus. This makes them less reliable for power-hungry or high-bandwidth devices like external SSDs, HD webcams, external GPUs, and multiple monitors.
Kepler Computing, a seven-year-old San Jose startup, has emerged from stealth mode claiming a novel 3D-stacking method and proprietary material can boost memory chip density without relying on costly EUV lithography. The approach applies to both high-bandwidth memory used in AI systems and SRAM cache found in CPUs and GPUs, and can reportedly be built using existing fabrication plants rather than requiring new facilities.
Samsung Electronics showed mockups of two new 3D memory designs, zHBM and zNAND-O, at the FMS 2026 conference in California. zHBM vertically stacks high-bandwidth memory directly on top of AI accelerator chips along the z-axis, cutting the distance data travels between memory and processor compared to today's side-by-side layouts.
Microsoft has launched Project Zenith, a stripped-down version of Windows 11 aimed at AI developers, requiring at least 64GB of unified memory and 250GB/s memory bandwidth. It will debut on AMD's Ryzen AI Halo hardware, a rival to Nvidia's DGX Spark, before expanding to other OEM devices, and ships preloaded with tools like VS Code, GitHub Copilot, Python, WSL, and .NET 10 already configured for coding.
A hobbyist bought two ASRock BC-250 mining cards from AliExpress for £126 and £103 respectively, and successfully unlocked all 40 compute units on both despite AMD's binning meant to disable some cores. One card also unlocked all 8 CPU cores, while the other had one faulty core, but both performed well running the gpt-oss-20b language model via llama.cpp, hitting up to 100 tokens per second at 8K context.
Apple's newly unveiled Mac mini with M6 and M5 Pro chips carries several specs the company didn't spotlight at launch. Memory bandwidth actually varies by RAM configuration, topping out only with the M5 Pro model, and the machine includes an Apple-designed N1 chip enabling Thread smart-home support. Additionally, units will ship with the unreleased macOS Golden Gate already installed.
SK Hynix held a groundbreaking ceremony at Purdue University in West Lafayette, Indiana, for its first U.S. facility, a $4 billion plant dedicated to packaging high-bandwidth memory chips. CEO Kwak Noh-Jung said the site will become a key American production base for HBM by 2030, with its first cleanroom expected to open in October 2028, though actual chip manufacturing will still occur in South Korea and China.
At Hot Chips 2026, researchers Anurag Agarwal and Radhakrishna Giduthuri presented a tutorial on High Bandwidth Flash (HBF), a package-integrated flash memory technology built like HBM but functioning more like an SSD. Since no HBF hardware exists yet, the talk relied on simulations and projections, examining how software—especially inference frameworks like vLLM—would need to be redesigned to exploit HBF's large capacity despite its coarse-grained, storage-like access pattern.
GoKawiil