Few direct matches — filled in with the latest updates.
Tag: #fp83 results
Meta/PyTorch engineer details MXFP8 GEMM kernel design in CUDA + PTX, achieving up to 99% of cuBLAS performance despite FP8 constraints. Includes deep dives into challenges and links to PyTorch integrations for DeepSeek-V3 training on B200 GPUs.
Moore Threads achieves systemic FP8 breakthrough, syncing with Nvidia globally as one of few domestic GPU makers. MTT S5000 delivers 1000 TFLOPS FP8 AI compute with 80GB VRAM and full precision support. MUSA platform enables compatibility with PyTorch, vLLM, and other AI frameworks.
Feather emulates FP8 inference on Ampere GPUs like RTX 3050 using custom Triton kernels and bit-packing, achieving 1.5x speedup over FP32 for TinyLlama with minimal accuracy loss. It targets memory bandwidth optimization without native hardware support. Future plans include CUDA Graphs, block quantization, and Llama support; accepted at PyTorch Conference Europe 2026.
Patients are increasingly bringing answers from Doubao and other AI assistants into hospitals, sometimes challenging doctors' diagnoses and requesting unnecessary tests. The article documents how AI hallucinations, incomplete symptom interpretation, and misplaced patient trust are increasing consultation time and worsening doctor-patient friction.
Unitree has launched a biomimetic seven-axis dexterous robotic arm, priced from 9,900 yuan. The product could lower the hardware cost of embodied AI development, robotic manipulation research, and automation prototyping.
Kling AI generated more than RMB 850 million in Q2 revenue, up 240% year over year, making it the standout growth driver in Kuaishou’s otherwise slowing business. Kuaishou is prioritizing AI investment, spinning Kling AI out for independent financing at an implied valuation of US$18 billion despite near-term profit pressure.
Rapidus is pursuing mass production of advanced 2nm semiconductors through a large-scale Japanese national project. Instead of matching TSMC’s scale, the company plans to compete with RUMS, an integrated one-building production model designed for diverse, lower-volume manufacturing.
Shuto Expressway struggled to expand generative AI adoption because employees were unsure what to use it for and the organization prioritized safety. By using Google Gemini and Gemini Notebook as a practical starting point, the company reportedly increased the number of users tenfold, including among senior employees.
The article investigates how ChatGPT answers questions such as “What do you recommend?” It suggests that ChatGPT may begin with a preselected shortlist of brands or products before conducting a search.
Google is adding an AI chatbot interface that lets users describe the content they want in their Discover feed. The feature will automatically adjust feed recommendations and remember preferences for future visits.