News for hackers. Stuff that matters.
Re:Google TPU Performance Analysis by buildbot on Wednesday August 23, 2017 @05:03PM (#15085110)
I would argue ECC is completely pointless during neural network inference, the amount of change a single bit flip, even a sign change, is likely to generate is minimal.
Re:Google TPU Performance Analysis by vomjom on Wednesday August 23, 2017 @10:39PM (#15087207)
An exponent bit flip, especially at later layers, would completely break inference.
Re:Google TPU Performance Analysis by puzzle on Wednesday August 23, 2017 @11:09PM (#15087335)
I'd be concerned about code, too. It's not unheard of for a single bit flip to make petabytes go away (true story).