3 ms·
8-Bit Numerical Formats for Deep Neural Networks (2022)
- ashvardanian 3y agoI am curious if there is anyone familiar with the standardization progress of those 8-bit representations? PS: Assuming they can only have 256 unique values, would have been great if the papers provided conversion tables for every 8-bit float to a 32-bit value, to make testing accuracy degradation easier across applications.