nvidia

15
Jump in the discussion.

No email address required.

FP16 -> FP8 -> FP4

Who made this shit graph?


:#marseyviewerstaretalking:

Jump in the discussion.

No email address required.

next generation :marseyreactor: everything will be 1 bit floats

you can have zero or infinity :marseylovegigaorgy:

nothing will get done but it will get done faster than ever

Jump in the discussion.

No email address required.

wtf even are 4-bit floats

like do they jettison the sign bit, or do they have 1 sign bit, 1 exponent bit, and 2 mantissa bits?

https://en.wikipedia.org/wiki/Minifloat

The smallest possible float size that follows all IEEE principles, including normalized numbers, subnormal numbers, signed zero, signed infinity, and multiple NaN values, is a 4-bit float with 1-bit sign, 2-bit exponent, and 1-bit mantissa.

I guess I was slightly off. What's even the point of this?

Jump in the discussion.

No email address required.

Speed I guess, can process more fp4 ops with each gpu and if you can say you have a larger model cause all the weights have less precision


:#marseyviewerstaretalking:

Jump in the discussion.

No email address required.

https://i.rdrama.net/images/17195140760080988.webp

I guess I'm not an AI dev but this doesn't seem like enough values to be worth anything. Even 8-bit floats are pushing it but they have 16x as many distinct values (okay maybe not quite that much since there will be more NaN sequences but close).

Jump in the discussion.

No email address required.

I'm not either but they have some models out there with 1 bit weights that seem to work (poorly)


:#marseyviewerstaretalking:

Jump in the discussion.

No email address required.

:marseyblackcop:

Jump in the discussion.

No email address required.

Link copied to clipboard
Action successful!
Error, please refresh the page and try again.