Dan Boneh

Bit-exact recomputation requires an exact model of each device's matrix arithmetic, including rounding, subnormal numbers, accumulator width, and the grouping and order of summation.

For Nvidia GPUs, Bodash, Boneh, Komargodsky, and Srivastava, Pearl Research and Stanford built such a model called Hawkeye ML System twenty twenty-six.

Since Nvidia does not document these details, the model was reverse engineered by black box tests.

It reproduces tensor core products bit for bit on a CPU for the Ampere, Ada Lovelace, and Hopper generations, and the white paper adds Blackwell.

I was, I was an intern in college.

I was a, a TA for Dan Boneh, who teaches a lot of the cryptography and blockchain classes at, at Stanford.

And is a research advisor here.

Yes.