Nvidia's Project Digits is a 'personal AI supercomputer'

>> The IBM Roadrunner was the first supercomputer to reach one petaflop (1 quadrillion floating point operations per second, or FLOPS) on May 25, 2008.

$100M, 2.35MW, 6000 ft^2

>>Designed for AI researchers, data scientists, and students, Project Digits packs Nvidia’s new GB10 Grace Blackwell Superchip, which delivers up to a petaflop of computing performance for prototyping, fine-tuning, and running AI models.

$3000, 1kW, 0.5 ft^2

Digits is petaflops of FP4, roadrunner is petaflops of FP32. So at least a factor of 8 difference, but in practice much more. (IE I strongly doubt digits can do 1/8th petaflop of FP32)

Beyond that, the factors seem reasonable for 2 decades?

Why even use a floating point if you have only 4 bits? Models with INT8 features are not unheard of.

1 sign and 3 exponent bits. AFAIK at the small number of bits it’s basically a teeny tiny look up table so you can precompute the table to be whatever math you want. Having an exponent instead of mantissa just means that the values that can be expressed are not linearly separated.