Here’s a report <https://www.nextplatform.com/2020/01/27/european-processor-initiative-readies-prototype/> on a Europe-wide effort to produce a processor chip for an “exascale” supercomputer. It will have an ARM CPU core, with a RISC-V-based accelerator and a few other special-purpose units on the side. I was intrigued to see mention of an “AI-inspired bfloat16 format”. Floating-point numbers commonly come in 32-bit and 64-bit sizes. There is also a “half-precision” 16-bit format, with reduced precision and dynamic range, commonly used in computer graphics, for example. This new format keeps the full dynamic range of 32-bit floats, at the expense of even further reduced precision. Apparently this works better for various machine-learning algorithms.