Rust wgpu now has a
-
Rust wgpu now has a .agents/skills, a .claude/skills, an AGENTS.md, and a CLAUDE.md.
Increasingly wondering if my goal of living in a way that I never benefit from the actions of generative AI users and generative AI users never benefit from the actions of me is going to inherently mean never using hardware-accelerated video on a computer again.
@mcc For devil's advocacy, simulated annealing is a (very simple) form of generative AI, and it has been used for semi-automated chip layout design for decades.
-
@amarioguy @ireneista @mcc @MaddieM4 oh the field solver (openEMS) has been around for longer than that, the problem has been that it was impossible to use unless you were also a grad student working in that space
@r @ireneista @mcc @MaddieM4 right right, the point is more "there is active work being done in the open silicon space that definitely wasn't happening in the last decade or so"
-
@MaddieM4 @mcc sadly i suspect for most casual people (not nerds like us, like the average person you'd see elsewhere and to be very clear i am VERY supportive of this goal of a proper open sourced GPU) unless it's at feature parity with NVIDIA (which i hear those kinds of drivers and underlying firmware/ISAs are nightmares to develop and maintain) i doubt much outside nerd-sphere would care enough to even gain significance
that + the fact that i also hear it's apparently ruinously expensive to develop GPUs (like everything about them i hear is only getting more expensive by the day, especially on the performance side) that a serious one would likely have serious maintainability issues i fear
@amarioguy Much of the ruinous expense comes from ordinary chips being write-only, and modern GPU design being immensely complicated. FPGAs suffer from these problems a lot less. In foreseeable future, we might even see FPGAs that bump their manufacturing yield up by dynamically routing around a couple of bad logic cells on the die.
-
@ireneista @amarioguy @mcc @MaddieM4 ohhhh we are very interested in this, i think you're the first we've seen to *name* this
we've definitely been around the block with hellprojects such as "how much can you overengineer a pub-sub system while still having your robot run?"
-
@ireneista @decay @amarioguy @MaddieM4 i suspect vector units are widely underutilized, and thus skippable, especially if you have access to a coprocessor that can run compute shaders. maybe as compilers have got better this (vector units are underutilized) has become less true.
@mcc FWIW, I've been toying with things that do Unicode on SIMD. The utilisation has definitely been going up.
-
@ireneista @amarioguy @mcc @MaddieM4 ohhhh we are very interested in this, i think you're the first we've seen to *name* this
we've definitely been around the block with hellprojects such as "how much can you overengineer a pub-sub system while still having your robot run?"
@ireneista @amarioguy @mcc @MaddieM4 there's also "that" fedi post making the rounds recently-ish pointing how GIMP fits into this type of space in an "interesting" way
-
@ireneista @amarioguy @mcc @MaddieM4 there's also "that" fedi post making the rounds recently-ish pointing how GIMP fits into this type of space in an "interesting" way
@r @ireneista @amarioguy @mcc I have not seen that one, and I'm curious
-
@r @ireneista @amarioguy @mcc I have not seen that one, and I'm curious
-
@ireneista @mcc @decay @amarioguy That's a good particular example. A part of computing freedom that I think is pretty important is "being able to make music with computers, ideally in realtime." I have that on my list of vital user stories for my someday project microkernel OS. The arts are the kind of thing that make computers worth doing.
A lot of realtime music plugin stuff depends of Fourier transforms and SIMD in general.
@MaddieM4 Well, actually, music is so slow that a multi-gigahertz processor could probably do enough DSP for real-time music processing just on its fancy FPU.
-
@MaddieM4 Well, actually, music is so slow that a multi-gigahertz processor could probably do enough DSP for real-time music processing just on its fancy FPU.
@riley @MaddieM4 @ireneista @decay @amarioguy To be clear I'd really like to be using vector units for my music software. But this whole conversation is about compromising on things.
-
@mcc For devil's advocacy, simulated annealing is a (very simple) form of generative AI, and it has been used for semi-automated chip layout design for decades.
@riley "generative AI" is an attempt to describe not a technology but a certain type of social/business practice developed after 2018 and exemplified by the company named "OpenAI". it doesn't have much to do with AI or generation. We call it that because in order to interact with other humans we must designate signifiers to denote referents and this designation is a mutual consensus process.
-
@aparrish @mntmn @mcc Nightmare scenarios I didn't even think about LOL. IME, the vast majority of ppl in FOSS FPGA toolchain-land despise AI. There have been small programmable logic chips taped out on TinyTapeout. I don't imagine the old chips will be changed from the final AI-less designs b/c that Costs Money. I'm personally going to cross that bridge when we come to it, for my own sake :D.
-
@ireneista @mcc @decay @amarioguy That's a good particular example. A part of computing freedom that I think is pretty important is "being able to make music with computers, ideally in realtime." I have that on my list of vital user stories for my someday project microkernel OS. The arts are the kind of thing that make computers worth doing.
A lot of realtime music plugin stuff depends of Fourier transforms and SIMD in general.
@MaddieM4 @ireneista @mcc @decay @amarioguy i see this (digital music) a lot (even though i don't really get it myself), which is where i feel the need to bring up that:
imo our entire pedagogy for "signals and systems" needs to be re-done or at least significantly revised
i personally survived the "cabal of Elite
️ universities" style of teaching it, but i've since encountered countless people who have given up on it because this style doesn't work for them -
@aparrish @mntmn @mcc Nightmare scenarios I didn't even think about LOL. IME, the vast majority of ppl in FOSS FPGA toolchain-land despise AI. There have been small programmable logic chips taped out on TinyTapeout. I don't imagine the old chips will be changed from the final AI-less designs b/c that Costs Money. I'm personally going to cross that bridge when we come to it, for my own sake :D.
@cr1901 @aparrish @mntmn it does seem like periods of many, many years pass between the popular FPGA chips changing in literally any substantial way at all. this said, wq was talking a few weeks ago about how atypical changes in "preferred" commodity FPGA chips *are* actually happening kind of rapidly right now due to "wars are happening in the vague zone between Yemen and Estonia" related issues
-
@cr1901 @aparrish @mntmn it does seem like periods of many, many years pass between the popular FPGA chips changing in literally any substantial way at all. this said, wq was talking a few weeks ago about how atypical changes in "preferred" commodity FPGA chips *are* actually happening kind of rapidly right now due to "wars are happening in the vague zone between Yemen and Estonia" related issues
-
-
@ireneista @r @MaddieM4 @mcc @decay @amarioguy they don't bother teaching you why it's useful. https://mastodon.social/@hyc/116947095099572153
-
@decay @riley @MaddieM4 @ireneista @amarioguy That's very interesting, but presumably the reason GPU companies don't do this already is because committing to a fixed ISA makes it much harder to rev architectures.
But it is the case that if I attempted an OSS GPU, in step one the "shader units" would almost certainly just be compiling the shader programs to RISCV plus one of the vector extensions, grabbing a RISCV softcore, and splatting twenty of them on the fabric.
-
@ireneista @r @MaddieM4 @mcc @decay @amarioguy they don't bother teaching you why it's useful. https://mastodon.social/@hyc/116947095099572153
@hyc @ireneista @r @mcc @decay @amarioguy It's such a common error. If anything you should be LEADING with the why.
My pet conspiracy theory why they don't, is that most curricula are designed by committee and lean toward including everything that sounds like it might be useful. This lack of discretion means that most of the material IS pointless for the students, and "well you might need it some way" is cover for "we can't explain why because there IS no good answer behind the curtain." And in situations where there IS a good answer, there's so much bad faith precedent that the students can't tell the difference.
This is why I think project-centered education would be revolutionary. It leads you to discover things in contextualized, justified, concrete ways.
-
@ireneista @MaddieM4 @mcc @decay @amarioguy the "traditional" curriculum (or at least, mostly generalizing from my own experience) is approximately:
1. some general engineering math (complex exponentials, linear algebra)
2. fourier series, fourier transforms
3. convolution
4. sampling
5. discrete time fourier transforms, discrete fourier transforms
6. laplace transforms, z transforms
7. filters, "various properties" (pole/zero, causality, stability, bandwidth, etc.), "various design tools" (bode plots, root locus analysis, known/established filter designs, etc.)
8. multiple-input and multiple-output systems
9. newer "state space" techniques
10. linear-time-varying systems, nonlinear systems, stochastic systems, and beyondwith "digital signal processing" as an offshoot that lives somewhere vaguely around step 7 with some ideas from later steps
this is..... an *extremely* theory-heavy and tbqh rather old (in terms of absolute-time) approach, very much within the "electrical engineering"-owned-and-controlled part of the conceptual space
