You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
FP8 Formats for Deep Learning
A group of NVIDIA, ARM, and Intel researchers got fp8 training working reliably, with only a tiny accuracy loss compared to fp16.
8-bit floating point (FP8) binary interchange format consisting of two encodings - E4M3 (4-bit exponent and 3-bit mantissa) and E5M2 (5-bit exponent and 2-bit mantissa). While E5M2 follows IEEE 754 conventions for representatio of special values, E4M3's dynamic range is extended by not representing infinities and having only one mantissa bit-pattern for NaNs. We demonstrate the efficacy of the FP8 format on a variety of image and language tasks, effectively matching the result quality achieved by 16-bit training sessions.
reducing the number of NaN/Inf encodings in fp1-4-3 down to just one bitstring
how much accuracy loss does this approach cause? They find that, across a huge array of models and tasks, the consistent answer is: not much—around 0-.3% accuracy/BLEU/perplexity:
The text was updated successfully, but these errors were encountered:
via https://dblalock.substack.com/p/2022-9-18-arxiv-roundup-reliable
The text was updated successfully, but these errors were encountered: