[HN Gopher] Transformer Inference Arithmetic (2022)
___________________________________________________________________
Transformer Inference Arithmetic (2022)
Author : lawrencechen
Score : 41 points
Date : 2023-05-13 08:17 UTC (2 days ago)
(HTM) web link (kipp.ly)
(TXT) w3m dump (kipp.ly)
| thatcherc wrote:
| Author seems to be using billion == 10^12 instead of the common
| billion == 10^9. A lot of the math still works out since there's
| a multiply and a divide by a billion, but it is a little
| confusing to see passages like this:
|
| > Given the parameter count, we can multiply by two to get bytes.
| So to calculate the size of the weights for a 52B model.
|
| > 52e12[?]2 = 104e12 bytes [?] 104GB
| pulvinar wrote:
| I assume that's a slip back to the old British billion. But
| luckily that mistake doesn't affect that result, and does get
| it right in the next equation:
|
| > 2,097,152 bytes [?] 0.002GB
| kipply wrote:
| oops will fix. never using names of numbers of zeroes ever
| again. that stuff sucks, i will express everything in
| scientific notation
| gxh8N wrote:
| Very nicely written. I also like how it changes color every time
| I reload the article.
___________________________________________________________________
(page generated 2023-05-15 23:02 UTC)