AI Briefing#gpu
The latest AI and developer news about #gpu, with the original source and a short summary.
The latest AI and developer news about #gpu, with the original source and a short summary.
Intel is preparing a next-generation Xe3P data center GPU equipped with 160GB of LPDDR5X memory to address the HBM shortage problem.
The sharp decline in AI inference costs is being driven by software, and local open models are quickly catching up to frontier model performance.
Meta is receiving $3.3 billion in tax breaks for its $10 billion AI data center under construction in Louisiana
Anthropic CFO Krishna Rao explained the company's compute and capital allocation strategy in his first podcast appearance.
AI agent demand is exposing the limits of GPU- and HBM-centric infrastructure.
Modal reduced serverless GPU startup to around 50 seconds using buffers and checkpoint restore.
The article explains how NCCL optimizes GPU collective communication on top of AWS EFA.
Nvidia has unveiled cuda-oxide, an experimental compiler that compiles Rust code directly to PTX, enabling the writing of safe GPU kernels.
The center of gravity for AI inference is moving from GPUs to the memory hierarchy.
Span, backed by Nvidia, is pushing to build distributed mini data centers that tap into home electricity.
The llama.cpp b9095 update enables tensor parallel processing on Blackwell PCIe GPUs without NCCL.
Mojo released its 1.0.0b1 beta, emphasizing Python interoperability and GPU strategy.
As Agentic AI adoption progresses, the ratio of CPUs to GPUs is changing, requiring new infrastructure designs.
Elon Musk's SpaceX is leasing a massive supercomputer to Anthropic, accelerating its push into the AI infrastructure market.
AMD has unveiled the MI350P PCIe AI GPU, which can be plugged into existing servers.
Anthropic has significantly expanded usage limits for the Claude API and Claude Code by securing large-scale GPU capacity through SpaceX's Colossus 1 infrastructure.
Anthropic has decided to use SpaceXAI's supercomputer resources to expand Claude service capacity.
Anthropic signed a compute capacity deal with SpaceX and raised usage limits for Claude Code and API request limits.
An analysis of Hugging Face user data found that AI builders tend to prioritize VRAM capacity over compute performance.
A summary of the principles and parallelization strategies for scaling LLM throughput near-linearly on TPUs and GPUs.
The RTX 6000 Pro was significantly faster than the 4080 Super at Qwen3.6 27B inference.
A build method for using CUDA and ROCm simultaneously with llama.cpp on Windows was shared.
On a single RTX 3090, Qwen3.6-27B achieved a 218K context and stable tool calling.
On an RTX 5080 16GB, Qwen3.6-35B-A3B hit 30 t/s at 128K and 89 t/s for new tokens.