Hacker News Reader

It's used inside Google's TPUs[0]. Useless for regular logic programming (You've wouldn't want a nuclear power plant, or your payment processor to approximately work), but has found a use case in scaling ML pipelines where the accuracy of individual values is less important compared to billions of parameters in aggregate.

[0] https://ieeexplore.ieee.org/document/9264836

View on HN

I don’t know that it’s “useless” for logic programming. I agree it’s less likely to be used for normal everyday stuff but it could see more proliferation (eg video decoding). I think TPUs are an interesting first step but the holy grail (which indeed may be impossible) is to be able to reduce large scale programs to run on approximate circuitry. For example, if chrome used a fraction of the memory, CPU, and power, that would be significant even though no one would notice that the page renders slightly imperfectly if executed well. As the paper notes TPUs are the first application although I think what we’re doing there is probably quite primitive be what the researchers in the area are working on long term.

Reply on news.ycombinator.com