Topic
Explore English articles about Generative AI.
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.
Windows 11のプレビュー版内部から、GPUやAI向けに統合メモリを手動配分する隠し機能が見つかった。RTX Sparkなど大容量統合メモリPCの登場を前に、OS側のメモリ管理はどう変わるのかを読み解く。
Canonical is investing in a three-year PhD research project to convert hundreds of thousands of lines of C into safe Rust. Using AppArmor and other real-world tools as test cases, the work will examine whether generated code preserves behavior and remains maintainable.
Microsoft Paint records a server-issued GUID in both the AI-generated image and its C2PA data, even for local generation. Photos was found to use a similar code path, but no personal identification or account matching has been confirmed.
Monthly payroll records covering 3.5 to 5 million U.S. workers from ADP show that employment in AI-exposed occupations for workers aged 22-25 failed to keep pace with growth in less-exposed occupations, with the gap showing up mainly in hiring rates. Because it's an observational study, causation remains unproven.
NVIDIA's Groq 3 LPX, now in mass production, splits AI inference work between GPUs and LPUs to reach 3,431 output tokens per second. Here's what that number means, where the benchmark falls short, and what's behind the $17 billion deal.
NVIDIA is reportedly discussing a follow-on investment in Perplexity at a valuation exceeding $30 billion. The talks come as Perplexity's agentic business grows and its ties with NVIDIA span GPUs, CPUs, and model development.
Xiaomi's second-generation SoC, the XRING O3, stays on 3nm while expanding to 10 high-performance CPU cores, a 16-core GPU, LPDDR6, and a 200TOPS NPU. This article separates official figures from measurements on a development board to examine what the 113.8GB/s memory bandwidth and 20W-class peak power actually mean.
d-Matrix's Raptor accelerator measured roughly 105 TB/s of card-level bandwidth using 3D-DRAM in real silicon. But its inference speedups come from architectural modeling, not third-party benchmarks, and production yield data remains unpublished.
A peer-reviewed study testing 10 AI models found that majority-following behavior in AI agents varies with group size and model. The 1,000-agent figure is not evidence of complex cooperation but the upper limit tested in a binary-choice experiment.
