Happy WebGPU Day
Yesterday was a big day for the Web: Chrome just shipped WebGPU without flags in the Beta for Version 113. Someone on Nomic’s GPT4All discord asked me to ELI5 what this means, so I’m going to cross-post it here—it’s more important than you’d think for both visualization and ML people. (thread)
So: GPUs are processors on basically every computer/phone. Individually they’re weaker than CPUs, but they run in packs of little ones that run in parallel. The G is for ‘graphics,’ but it’s turned out they’re good for anything involving lots of math–like ‘AI’, which at core boils down to lots (and lots and lots) of matrix multiplication operations. To do math, not graphics, on a GPU you need an API/language for them; the most important of these is CUDA, which is tightly coupled to NVidia and a real PITA to set up.
On the web, we’ve only been able to access the GPU through something called WebGL. It’s old, and while you can do some neat stuff with it, it’s fundamentally built for graphics, not for the matrix-multiplication type stuff that is the bread and butter of deep learning models. Since WebGL launched in 2011, lots of companies have been designing better languages that only run on their particular systems–Vulkan for Android, Metal for iOS, etc. These are great where they work, but even harder to run everywhere than CUDA.
WebGPU is an API and programming that sits on top of all these super low-level languages and allows people to write GPU code that runs on all of them–that is, on just about any phone/computer with a web browser. This is a big deal, because it has “compute shaders” that lets you write programs that take data and turn it into other data. Working with data in WebGL is really weird–you have to do things like draw to an invisible canvas and then read the colors as numbers. In WebGPU, you can just do math. Really fast.
That means it’s actually capable of doing–say–inference on a machine-learning model like GPT4All, multiplications on data frames, etc. There are already some crazy things out there, like a version of Stable Diffusion that runs in your web browser.
This will be great for deepscatter too. Maps like https://atlas.nomic.ai/map/twitter can render 5,000,000 tweets incredibly fast, but need a lot of CPU for compute. Often it’s fast enough, but real-time rendering needs to run 30x a second: I have a long and growing list of things that are nearly impossible in WebGL but will be quite easy in WebGPU.
Right now it’s only released on Chrome, but it’s not an only-Google thing forever. It’s an honest-to-goodness W3C standard like HTML, CSS, or SVG. All the browsers have been working on it; Chrome is just shipping first because Google is rich compared to Safari and Firefox. One of my favorite parts about reading the minutes of the WebGPU committee over the last year is watching people from the other browsers jealously grouse about how much money Google throws at Chrome.
JB: Corentin mentioned that all the browser vendors have been at the table, for a long time. Haven’t you had a long enough chance to give that feedback already? Answer is - no. :) Our impl isn’t done. Not about whether a certain period of time has elapsed - but rather do you have an impl that satisfies the criteria. Chrome’s one of the best funded orgs in KR: Without going too much into funding, thinking about spec criteria, we had a list of bugs triaged into v1 and post-v1. Let’s burn that down to zero, and if we consider larger change, we should probably let them sit as they are. There’s probably a way to implement something reasonable later. We can probably do these changes in a compat way in the future. Let’s get issues down to zero. Impl feedback is useful of course. We don’t go to rec without multiple impls. Looking at wording, I don’t think “canditate rec” is gated on mult implementations.
But they’ll come along–the Chrome-derived ones like Edge first, but Safari and Firefox eventually too because GPU compute is just such an important thing. And when they do, it rescrambles the whole compute stack. Slowly but surely real GPU compute, tensor operations, all the stuff that makes AI tick moves from something that happens only in the cloud, to something that can get reshuffled, rearranged, and done privately on PCs again. Another chance to reclaim compute from the cloud.