The military has not spearheaded the current AI revolution because modern LLM development depends on vast, unclassified civilian data, fast commercial capital cycles, and the dual-use nature of the te…
BreezeBlue released Breeze TTS 2, an open-weight bilingual text-to-speech model with sub-40ms time to first audio on an NVIDIA H100, 0.32 real-time factor streaming, and voice cloning, design, and dir…
A forum user on Level1Techs is building a 4U chassis to connect four air-cooled or eight liquid-cooled NVIDIA H100 GPUs to a Dell PowerEdge R7515 server via PCIe switches, using either a Broadcom PEX8…
The NanoGPT speedrun repository reports that a collaborative effort has trained a language model to 3.28 cross-entropy loss on the FineWeb validation set in under 75 seconds on 8 NVIDIA H100 GPUs, a d…
Liquid AI released DSpark draft model checkpoints for three LFM2.5 models, enabling speculative decoding that speeds up inference by up to 3.18x on GPUs and 2.87x on-device without changing output qua…
KernelArc, a multi-agent framework for autonomous GPU kernel optimization, achieved first-place rankings on representative L1, L2, Quantization, and FlashInfer tasks at the public SOL-ExecBench leader…
A looming AI debt bubble, driven by massive infrastructure spending on GPU clusters and data centers, threatens systemic financial stability, according to analysis of the tech sector's capital expendi…
A team from Nagoya University and the University of Tokyo used Claude Code to port CReSS, a 250,000-line Fortran typhoon simulator, to GPUs, achieving 162 validated GPU kernels and a 5.1x speedup on a…
A technical analysis by an unnamed author, inspired by a post from Su Jianlin, derives the arithmetic intensity (FLOPs per byte) of attention during single-token decode, finding MHA has an AI of exact…
Attestable, a zero-knowledge proof startup, announced it can prove production-scale transformer inference at 85 tokens per second on a single NVIDIA H100, with proof sizes of 4.35 to 7.92 MiB and veri…
Cast AI benchmark testing shows that continuous batching at batch size 8 reduces Llama 3.1 70B inference cost on a single H100 from approximately $0.60-$0.80 per million tokens to $0.15-$0.25 per mill…
Google DeepMind released DiffusionGemma, an experimental open-weight model that converts the Gemma 4 architecture into a discrete text-diffusion system using less than 10% of Gemma 4's original traini…
Google Cloud is collaborating with MLCommons through the MedPerf initiative to use Confidential Computing for privacy-first benchmarking of medical AI models, enabling validation on diverse real-world…
UiPath has built a high-performance GPU platform on Google Cloud's AI Hypercomputer to support its enterprise agentic AI initiatives, moving from isolated clusters to a shared fleet that balances A3 V…
Google DeepMind introduced DiffusionGemma, an experimental open-weight language model that uses discrete diffusion to generate text at high speed, achieving roughly 1,500 output tokens per second on a…
NVIDIA reports that two AI computing clusters built from identical H100, GB200 NVL72, or GB300 NVL72 systems can deliver 8% to 12% differences in training throughput due to configuration choices in ke…
A developer built MEM v3 (Model Execution Manager), an orchestration engine using Python, PyTorch, and DeepSpeed that sustains 1 million training steps with zero out-of-memory errors. The engine featu…
ANEForge, a new open-source Python package, compiles tensor graphs directly into programs for the Apple Neural Engine (ANE) and runs them without CoreML, enabling both inference and training on the ch…
Microsoft is replacing OpenAI's image generation models in PowerPoint and Bing with its own MAI-Image-2, the clearest sign yet that its $13 billion investment doesn't guarantee a permanent product pla…
A benchmark study comparing standard non-confidential execution with confidential computing mode on a single NVIDIA H100 80GB GPU hosted in an Intel TDX confidential instance found that confidential m…