This explains the cooperative relationship between Huawei and DeepSeek: Huawei supplies Ascend hardware and a native software stack while DeepSeek provides workloads, models and kernel engineering that make that hardware usable at scale. The release of Ascend support into DeepSeek repositories (TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect) embodies that trade, and the authors argue DeepSeek will primarily run on Huawei’s stack by 2027-2028 as China builds a domestic ecosystem. The piece traces how Huawei’s existing Ascend ecosystem and DeepSeek’s engineering complement each other, with mutual adaptations such as FlashMLA’s reorganized KV layout to reduce memory transfer engine (MTE) requests.
Technical findings center on an Ascend 950PR_9579 (28 cube cores, 128 GiB HBM) microbenchmark suite and software analysis. Huawei’s matmul kernels hit up to 99.7% of theoretical ceilings on BF16 GEMMs, while Ascend’s median achieved/attainable fraction on compute-bound shapes is 41.5% versus H100’s 46.8%; absolute throughput places 950PR between A100 and MI300X. Ascend shows strong small-message HBM utilization and an MXU design with L0A/L0B/L0C tiling that boosts reuse; a copy-method switch introduces a mid-curve bandwidth dip. Ascend variants are binned across compute and HBM due to yield. On-chip Linx816 AICPU cores handle short CPU-like tasks and indexing; TileLang’s Ascend backend and AutoSchedule yield up to 40% execution reductions on small sizes, while full CANN/HDK support remains in flight.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Simon Edwardsson trained a 360M-parameter language model to simulate a terminal, predicting screen changes based on keypresses without a real shell or filesystem. The model can create files, run programs like vim and tmux, and remember previous commands, demonstrating a new way to interact with AI in a browser. (simedw.com)
Claude now integrates with Google Docs, Sheets, and Slides through a new add-on and connectors in beta, allowing users to work directly within their files. It can read, edit, and suggest changes in documents, spreadsheets, and presentations without switching apps or tabs. (claude.com)
LLM agents were tested in a Pokémon Showdown 2v2 environment to evaluate their ability to coordinate and communicate during battles. Results showed that free communication improved win rates compared to restricted modes, demonstrating the potential for collaborative AI in strategic environments. (olivergrabner.com)
Linden is a low latency layer designed to verify AI outputs against schemas, data, and rules to ensure decision alignment with business requirements. Despite efforts to promote it, the developer has not gained any signups and questions whether the product serves a real purpose. (news.ycombinator.com)
Chopin is a real-time multiplayer environment designed to enhance team planning with AI agents, reducing cognitive load and improving decision-making. It emphasizes the importance of planning and review in software development, leveraging AI to streamline coding tasks and focus on strategic aspects. (githubnext.com)
OVHcloud, a French cloud computing company, aims to develop frontier AI models to become Europe's second large language model (LLM) provider by 2026. The company is investing in AI research to expand its presence in the AI industry. (reuters.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.