A new 16-bit floating point format for machine learning! https://hub.packtpub.com/why-intel-is-betting-on-bfloat16-to-be-a-game-changer-for-deep-learning-training-hint-range-trumps-precision/
It increases range at the expense of precision, and often allows 16-bit computation (smaller, faster hardware) to replace 32-bit ML logic.
miniblog.
Related Posts
Amazing blog post about running a modern Linux kernel on Chumby hardware, investigating bad CPU numbers from top, and finding a Linux kernel bug from 2009!
When someone says "you should use the cloud", I'm never sure if they mean:
1: use a virtual server rather than running your own hardware, or
2: use hosted services rather than running your own e.g. database
2 is often surprisingly expensive IME.
Remotely bricking satellite modems as part of (presumably) nation state hostilities: https://www.reversemode.com/2022/03/satcom-terminals-under-attack-in-europe.html
If over-the-air update systems aren't receiving regular patches, maybe it would be better for industry to use hardware that can't be reflashed remotely?