From: ThatWouldBeTelling@thevillage.com   
      
   BGB wrote:   
   > On 10/3/2025 11:40 AM, EricP wrote:   
   >>   
   >> The issue with FP8 support seems to be that everyone who wants it also   
   >> wants their own definition so no matter what you do, it will be unused.   
   >>   
   >   
   > As for FP8:   
   >   There are multiple formats in use:   
   >     S.E3.M4: Bias=7 (Quats / Unit Vectors)   
   >     S.E3.M4: Bias=8 (Audio)   
   >     S.E4.M3: Bias=7 (NN's)   
   >     E4.M4: Bias=7   (HDR images)   
      
   Its not just the memory formats, its also the operations.   
   In FP8 few may want to waste 1/8th of the encode space on NaN's.   
   Maybe not sticky infinity, rather saturate at max but not stick there.   
   Maybe no negative zero.   
      
   All of those encodings might be reallocated to values more useful   
   for that application.   
      
   They won't want to calculate single argument transcendentals like tan(x),   
   they will use 256 byte lookup tables.   
   The multi-operand functions ADD, SUB, MUL, would be faster in hardware   
   than 64kB lookup tables.   
      
   Also a lot of these are used matrix ops - super-duper-SIMD.   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)