↗
TONTOU Attack Bypasses Spectre v2 Defenses on AMD and Intel
MIT CSAIL researchers show a new attack, TONTOU, that uses timed interrupts to re-poison branch predictors and bypass Spectre v2 defenses on AMD and Intel CPUs.
Fresh signal from the moving edge of hardware, technology, and the systems underneath.
↗
MIT CSAIL researchers show a new attack, TONTOU, that uses timed interrupts to re-poison branch predictors and bypass Spectre v2 defenses on AMD and Intel CPUs.
↗
DeepSeek plans a significant increase in API prices as demand for V4-Flash surges, ending the era of ultra-cheap access for developers.
↗
Longsys is combining high-bandwidth memory, SSDs, and cache software to target the bandwidth and capacity limits of local AI systems.
↗
Olares will show its RTX 5090M-powered personal AI cloud at IFA Berlin, putting local models and user-owned data at the center of its European debut.
↗
Kioxia's GP1 Series PCIe 6.0 SSD reaches 10 million IOPS so flash can act as a GPU memory extension for AI, with samples due by year end.
↗
Anthropic confirmed it is building a custom silicon team to co-design chips and models, a step that reduces its reliance on Nvidia as Claude demand grows.
↗
SpaceX and Nvidia are designing Starmind AI1, a satellite compute payload with Rubin GPUs and Vera CPUs, as SpaceX commits exclusively to Nvidia.
↗
Nvidia is open-sourcing cuFile, the API that lets GPUs read and write storage directly, with Google, Intel, and Meta as founding maintainers.
↗
AI accelerators are running out of memory bandwidth before they run out of compute, and the whole stack is being rebuilt around it.
↗
Used and refurbished hardware is becoming a first choice for businesses and enthusiasts, and the industry is reorganizing around it.