[phh]

> How/when did these types of regulations come about?

I can't say much about US. As I see it, EU pretty much copied US about that part. There was nothing related to computation in the EU's AI Act projects until few months ago, it was purely a "what kind of data processing are you allowed to do?"

[alchemist1e9]

Politely, what the hell are you talking about? Who is telling anyone what they can or cannot compute?

[iamjackg]

US:

https://www.whitehouse.gov/briefing-room/presidential-action...

"Until such technical conditions are defined, the Secretary shall require compliance with these reporting requirements for:

          (i)   any model that was trained using a quantity of computing power greater than 1026 integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 1023 integer or floating-point operations[...]"

EU:

https://thefuturesociety.org/wp-content/uploads/2023/12/EU-A...

[geon]

> 1026

> 1023

Should be 10^26 and 10^23.

[alchemist1e9]

Probably I did this wrong but I’m getting an approximation of 300K H100s completes that in a month. At least they choose something fairly large it seems. Not sure how LoRA or other incremental training is handled.

[sbierwagen]

Depends on which spec you used, since the law doesn't specify the floating point width. If you used FP8 ops on the H100 SXM then a single GPU would hit the limit in 25265285497.72612 seconds. 300,000 GPUs would pass 10^26 FP8 ops in 23 hours.

[hayley-patton]

Are they trying to bring back SIMD-within-a-register? Though that only gives you ~one order of magnitude doing packed 4-bit stuff with 64-bit GPRs. And perhaps fixed-point, sign-exponent and posits are unregulated.

[cyanydeez]

anyone with a functional government.