Running NumPy in the browser now benefits from an accelerated BLAS because Emscripten-forge rebuilt a conda-like WebAssembly ecosystem and enabled a Fortran toolchain (Flang) and OpenBLAS builds that link to NumPy at runtime. That work required Flang patches to target wasm32, OpenBLAS adaptations for Emscripten calling conventions, and packaging so libopenblas is a separate, dynamically linked package. The language-agnostic distribution lets Python, R, and other stacks share the same optimized native libraries, so upgrading OpenBLAS improves the whole ecosystem without rebuilding NumPy. Key contributors include Isabel Paredes and Serge Guelton on Flang, Ian Thomas on OpenBLAS integration, and QuantStack on SIMD kernels.
The performance gains are substantial and specific. With OpenBLAS 0.3.34, 1024×1024 matrix multiply in float32 is about 30.92× faster (and 14.90× for float64) versus the no-BLAS baseline; Level‑3 GEMM shows the largest wins while some LAPACK-backed np.linalg calls see smaller 1.08-1.65× gains. OpenBLAS 0.3.35 adds wasm SIMD microkernels and yields another 1.79× (float32) and 1.41× (float64) improvement over 0.3.34 (e.g., 28→50 GFLOPS and 14.6→20.5 GFLOPS), and an optional Relaxed SIMD build boosts that further (about 1.17× float32 and 1.43× float64 on supported engines). Matrix-vector and vector kernels also improve in 0.3.35; remaining limits stem from LAPACK routines not yet specialized for WebAssembly.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Fitting a logistic curve from early data is nearly impossible because small errors in the data can cause large inaccuracies in estimating the limiting value when data is only from one side of the curve. This difficulty arises because the logistic function behaves almost exponentially in the tails, making the parameters highly sensitive to data errors. (johndcook.com)
Discrete-time Fourier series and transform are mathematical tools used to analyze periodic signals in digital signal processing. They decompose signals into their frequency components, enabling analysis and manipulation of digital data. (eli.thegreenplace.net)
Tiny radio transmitters revealed that white-backed swallows in Australia use deep torpor to survive extreme temperatures. This behavior was previously thought impossible for small passerines in such harsh environments. (theconversation.com)
Funding cuts by the Trump administration may prevent over 600 aging Americans from receiving the results of an Alzheimer’s disease study. The shutdown is part of a broader disruption affecting more than 5,700 medical research grants over two years. (theguardian.com)
Learning a new language involves complex mental processes that can help maintain cognitive health as people age. Research indicates that bilingualism and language acquisition provide cognitive benefits and enhance social connections for older adults. (theconversation.com)
GPT-6 Astra has decoded a WWI German radio cipher encrypted with the ADFGVX method, revealing a message about a British cruiser arriving at Sevastopol. The decoding involved using a specific encryption key and a detailed rearrangement process based on the key word 'TRUPPENVERSCHIEBUNG'. (prinzai.com)
Today's best Hacker News stories, summarized and screenshotted, one email a day.