Skip navigation
Engineers used to spend weeks optimizing chip layouts like this by hand.
0:00 / 0:22

Engineers used to spend weeks optimizing chip layouts like this by hand.

Rina

@irinatoxi3.8K followers

Visit on X
732 likesDownloadOriginal post

58,733 views

Engineers used to spend weeks optimizing chip layouts like this by hand. Here an algorithm does it live, trying 240,446 tiny moves, shifting, swapping and rotating blocks, and keeping only the ones that make the wires shorter. Total wiring drops from 2.18 m to 1.27 m, and all 368 jammed wires get untangled. Same chip and same parts, nothing added, yet it ends up using 27% less power, and signals that used to arrive late now arrive on time. More and more, this job goes to AI. Google used its AlphaChip to place blocks on its own TPUs, and the biggest chip design companies are shipping AI tools for exactly this. This one is a demo chip, but the trick is very real, and it’s already shaping the chips that run today’s AI.

16 Comments

Comments from the original post on X (Twitter)

  • Evin's profile picture

    @evinh111

    This is so insanely oversimplified lmao How do you define reductions in power and timing errors in this simulation? Does this account for heat? What is the node size? How are you assessing crosstalk and error rates? Is this actually manufacturable with yields?

    6 likes

  • Rina's profile picture

    @irinatoxi

    power and timing are just rough estimates from wire length, no heat, crosstalk or yield. it's a demo of the idea, not a real tapeout

    3 likes

  • Evin's profile picture

    @evinh111

    So this algorithm isn't actually "optimising a chip layout" at all. It's just moving some squares around and claiming numbers that don't mean anything. Cool cool.

    1 like

  • Rina's profile picture

    @irinatoxi

    the wirelength and jam numbers are real, measured on the layout. power and timing are estimates on top of that. fair point that I should’ve said that clearer in the post tho

    1 like

  • 琥珀青葉@KohakuLab's profile picture

    @KBlueleaf

    May be related:

    6 likes

  • Wayne Nilsen's profile picture

    @waynenilsen

    How does it avoid local optima? Perhaps rerun with different initial condition a few times.? Maybe there are just so few valid configs it doesn’t matter

    5 likes

  • Rina's profile picture

    @irinatoxi

    it takes some bad moves early on so it doesn't get stuck. and yeah reruns def help

    5 likes

  • cyo's profile picture

    @cyothevile

    Ok now wire the clocktree 🤣

    3 likes

  • nihasilev's profile picture

    @nihasilev

    Yani pahalı ram-işlemci-ekran kartları; yakın geleceğin çöpleri olduğu için Son bir vurgun yapıyorlar, gelecek teknolojiyi piyasaya sürmeden önce...

  • Metaforest's profile picture

    @Metaforest

    We do this in our heads without much thought once all the devices are defined. You make a mountain out of a mole hill!

  • Zayden Jung's profile picture

    @ZaydenJung

    Genetic Algorithm?

  • Enric Pastor's profile picture

    @EnricPastor4

    This is placement, channel creation and detailed routing. There are algorithms out there since the 80 or even earlier.

  • Luís Outeiro's profile picture

    @luis_outeiro

    Share? 😇

  • OMBI.MrPresentable's profile picture

    @CouncilShiba

    I do this on E frontier all day. I’m a ship simulator

  • John Hu's profile picture

    @Johnascos

    impressive

  • Soumil Bhandari's profile picture

    @SoumilBhandari

    240k moves live is crazy. as an ece student this is so cool to watch