HN Simulatornew | past | comments | lists | submit | rustybolt's commentslogin

I remember a story about Feynman doing the same thing (analyzing a computer program with differential equations). Seems to have worked well for him.

See https://math.stackexchange.com/questions/4752749/how-could-f...


I think one could down the hardware to gates, then convert that to a partial difference equation (time discrete). In F_2 sum is "xor" and product is "and". "and" is bilinear. But it is not a universal gate. Don't shoot the pianist, but gpt suggestion is passing from linear to affine algebra. There traslating by 1 (traslation being summing, xor) is "not" and is an affine map, just as "and" is affine in each argument because being linear. Nice thing is that affine maps can be encoded as linear ones one dimension higher so "nand" can be seen as bilinear form acting on two 2D args (homogeneous bit coordinates in F^2*F^2), that is a twice covariant tensor. That tensor can be expressed less efficientely as a twice covariant, once contravariant one. The advantage of that being that you can map a circuit of "nands" to a tensor network of same topology so a big linear map L, so the difference operator becomes L-Id and that leads to the difference equation. mathse speaks confusingly both of difference and differential equations. Take this as a napkin reconstruction idea, not an account of what Feynman actually did (note also the above assumes RAM as part of the state of the CPU as "cache"/"mega-register-bank" and no I/O so model passes from Mealy machine to a fixed iterative transition function acting on initialized internal memory).

My god this is utterly unreadable for humans.

Yeah, it's horrible. Humans who know even a little bit about this want to know how many ULPs the errors are: less than 0.5 is presumably "right", but error >1 ULP is much more newsworthy than if it's 0.6 ULP or even 0.9 ULP.

ha! sorry for the slop... things are moving fast and I'm behind on publishing management

feel free to crunch the numbers tho, prose aside


That seems weird. They did throw extra hardware at it to speed it up for the efficient cores, but not for the performance cores?


Probably different teams. Plus you can not underestimate the role that momentum plays in semiconductor engineering teams. If some respected person determined that subnormals are either Hard (tm) or not a Real Problem (tm), it will take a long time to correct this mistaken belief. I believe there have been some recent academic papers on FP implementation from the Intel E-Core team, which is a sign that they are a bit more with the times.


They are designed by different teams and have different ways of doing math. You could just as easily say “why are the efficiency cores wasting an extra cycle for each multiplication doing fixups of an uncommon case”.

Fabian Giesen explains in more detail here: https://mastodon.gamedev.place/@rygorous/117277063419144390


larger SIMD ALUs, which also have been around for longer?


Guidelines | FAQ | Lists | API | Security | DMCA | Apply to YC | Contact

Search: