Category
SpaceX plans to launch Vera Rubin-based AI satellites in Q4 2027. We examine what differs from ground-based racks, the average 175kW heat dissipation challenge, and where FCC approval stands.
Perplexity has released Portable Computer, which moves its AI agent's execution engine onto NVIDIA DGX Spark. Sensitive data stays on the device, while searches and harder reasoning tasks are handed off to the cloud only with permission.
SiFive's BigSky packs a 32-core RVA23-compliant RISC-V system into a 2U rack unit, aiming to accelerate OS and CUDA porting work. With limited availability, pricing, power consumption, real-world performance, and RHEL support status remain open questions.
SpaceX's planned $100 billion Starship spaceport on Louisiana's coast is examined through its 10-launch-pad design, tax incentives, local payments, and the FAA environmental review still ahead.
Nikkei Asia reports that China has been quietly delaying customs clearance for germanium and quartz-based materials shipped to Taiwan since 2025, without any public notice. Some shipments have been delayed by months, and some Taiwanese suppliers have lost orders. This piece examines the tactic of "silent restriction," which is hard to challenge legally or route around, and explores how the resulting demand could shift toward Japanese quartz glass makers.
Intel's Crescent Island pairs 32 Xe cores with up to 480GB of LPDDR5X in a 350W air-cooled PCIe card aimed at long-context inference. Memory bandwidth and real-world performance remain undisclosed.
Qualcomm has previewed its next-generation Oryon CPU running at 5GHz, along with FlexCache, a new technology that dynamically shares cache capacity across different core types. The peak clock speed is about 5.5% higher than the current official maximum, leaving sustained performance and power consumption as key questions still to be verified.
NVIDIA has released DLSS 4.5 Ray Reconstruction, built on a second-generation Transformer model, in early access for all RTX GPUs. We examine the override for 30 supported titles, scene-dependent image quality gains, and the limitations that remain.
Apple's M6 uses a 2nm process and three types of CPU cores to update its smaller Macs, while the M5 Ultra targets large-scale workloads with four third-generation 3nm dies and 1.2TB/s of memory bandwidth. We compare the official performance figures against previous generations and lay out the design differences and caveats.
OpenAI's first custom inference ASIC, Jalapeño, outperformed Nvidia's GB300 in fixed-length benchmark tests. This article examines its power-efficiency gains and the unresolved questions around Rubin comparisons, long-context validation, and mass production.
