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.
As enterprises overprovision GPUs, actual utilization is staying at around the 5% level.
An Nvidia vice president said AI costs now exceed labor costs.
With a 2x RTX 3090 Ti setup, applying `-sm tensor` increased llama-bench throughput.
A VRAM estimation tool that runs llama.cpp's fit algorithm in the browser has been released.
The analysis finds that demand for AI code agents has exploded, but infrastructure hasn't kept pace.
On Intel B70, llama.cpp OpenVINO became faster than SYCL, but still lagged behind LLM-Scaler.
NVIDIA unveiled Blackwell Day-0 support for DeepSeek V4 with performance of 3,500TPS per second.
A dual RTX 6000 build passed a long-term bench test under 1650W wall power and 95°C CPU conditions.
Qwen3.6-27B ran at 85–85–100 t/s on a 24GB RTX 5090 Laptop GPU.
The top score in the public inference challenge for Tesla T4·Qwen2.5-0.5B is 3,536 tok/s.
JWST and GPU/AI are enabling large-scale analysis of galaxies in the early universe.
It's a public catalog that updates GPU prices from 20+ clouds every 7 hours.
NVIDIA AI is delivering results in 5 fields including climate prediction, wildlife conservation, and recycling automation.
This is a benchmark result running MiniMax M2.7 4bit on dual Asus GX10 at about 100W.
On an RTX 3090, Qwen3.5-27B with DFlash+DDTree recorded up to 207 tok/s.
Morgan Stanley analyzed that the rise of Agentic AI will expand AI chip spending beyond GPUs to CPUs.
AI investment and performance have surged, but employment and public perception remain mixed.
Boosted Qwen3.6-35B-A3B speed by 54% on an RTX 5070 Ti using --n-cpu-moe.
The Intel Arc Pro B70 was compared with NVIDIA RTX and AMD Radeon AI PRO on Linux.
Unweight reduces LLM weights by 15~22% on H100 while keeping outputs bit-exact.
The UK government has launched a $675M AI fund to reduce dependence on US Big Tech.
xAI will provide Cursor with tens of thousands of GPUs to help train its Composer 2.5 model.
llama.cpp's ggml added graph_reused to reduce CUDA's O(n) change checks.