5 ms·

"640K ought to be enough for anybody." - etc etc etc. It's becoming increasingly clear that many of these techniques that for the past 10 years that have been

by giobox 4y ago

"640K ought to be enough for anybody." - etc etc etc. It's becoming increasingly clear that many of these techniques that for the past 10 years that have been derided "pointless" or novelty now have real applications. For one example, the automotive industry - ignore the hype on autonomy, computer vision is already delivering real benefits for active safety systems. I use github copilot every day - its not perfect but good enough to add value to my workflow. Apple's automated tagging of my photo library via computer vision allows me to discover hundreds of images of my life and family I'd forgotten all about. Stable diffusion can clearly replace an artist in some cases, ignoring the moral/ethical issues. I'm extremely excited for future of all this, frankly. The first step into such a new paradigm is always hard - people made the exact same "home computers are pointless" arguments in the late 70s/early 80s. I don't think anyone agrees with that anymore...

ripe 4y ago
> I use github copilot every day - its not perfect but good enough to add value to my workflow. Could you please share an example? I saw a description of copilot but couldn't imagine what it might be useful for.
paulmd 4y ago
Neural processing is ubiquitous in phones (all the flagships do neural enhancement on the image, and iirc apple does gesture recognition neurally), and all Ice Lake and newer laptop chips (plus lol Rocket Lake), as well as all Zen4 chips have neural instructions too. NVIDIA's had it for 2 gens now. It's pretty well across the stack at this point. And it's getting used for different things in different places. Things like face recognition or gesture recognition are natural fits in a low-power environment, reduces power consumption of those features hugely. PCs and servers can do content generation or some other larger "inference" tasks. Neural game upscaling is a huge win too. The neural-TAAU upscalers (DLSS 2.x and XeSS) perform a lot better than FSR still, even compared to FSR 2.0/2.1. And there will be other things they can figure out how to ML-accelerate as adoption continues, I'm sure. It enables some solutions to hard problems that don't have good deterministic algorithms, and you can run a lot bigger models than you can without acceleration. also fabrice bellard (of course) wrote a neural compressor... https://bellard.org/nncp/ https://bellard.org/nncp/ it's also likely going to be useful in level design and asset creation as well, although of course as we're seeing now with Stable Diffusion there's some interesting legal questions with content creation. optical flow engine (not neural) is another big win for computer vision stuff, too, offloads a few tough tasks to hardware, like object tracking and motion estimation. hooking into that with zoneminder or something would be awesome. I'm also stoked for shader execution reordering too. This is similar to what Intel calls "ray binning", it basically is a best-effort "re-alignment" of the thread (state,etc) to the most similar execution group for memory read alignment/coalescing. I think one or another raytracing implementation (Intel or NVIDIA) they were coalescing based on material. Intel talks about theirs here: https://www.youtube.com/watch?v=SA1yvWs3lHU https://www.youtube.com/watch?v=SA1yvWs3lHU Ada whitepaper: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvid... I am interested to hear what AMD is doing with RDNA3, supposedly there is a new RDNA3 ISA instruction for matrix acceleration but rumors are it's not a full high-performance matrix unit like CDNA and >=Turing? I don't know why you'd add an instruction without some hardware acceleration though. So maybe less than the other implementations... which is a little disappointing.
wyldfire 4y ago
> I use github copilot every day - its not perfect but good enough to add value to my workflow. Wow I thought this copilot thing was just like a joke. People use it to write software? I'm really curious, now. Can you share examples of the code you're writing with it?
- giobox 4y ago
  You can use it for just about anything that is text - it will complete configuration files, not just code, and even does a reasonable job a lot of the time writing method docs automatically. Thinking of it as just for code is already a narrow application. Sharing code examples is also of little relevance - just try it yourself and see if it fits your needs. The dollar/value ratio maps for my needs to avoid trips to docs or search engines for syntax confirmation - it may or may not for you. If it gets it wrong sometimes I don't really care, all the time it takes is pressing tab to complete or not to complete if suggestion is garbage. It's just an extension to the tab key's functionality, which is why it's so easy to use every single day when it integrates to most text tools. I also think if this is what we can have today, the future of code completion tools is very exciting.

by giobox 4y ago

ripe 4y ago

> I use github copilot every day - its not perfect but good enough to add value to my workflow. Could you please share an example? I saw a description of copilot but couldn't imagine what it might be useful for.

paulmd 4y ago

Neural processing is ubiquitous in phones (all the flagships do neural enhancement on the image, and iirc apple does gesture recognition neurally), and all Ice Lake and newer laptop chips (plus lol Rocket Lake), as well as all Zen4 chips have neural instructions too. NVIDIA's had it for 2 gens now. It's pretty well across the stack at this point. And it's getting used for different things in different places. Things like face recognition or gesture recognition are natural fits in a low-power environment, reduces power consumption of those features hugely. PCs and servers can do content generation or some other larger "inference" tasks. Neural game upscaling is a huge win too. The neural-TAAU upscalers (DLSS 2.x and XeSS) perform a lot better than FSR still, even compared to FSR 2.0/2.1. And there will be other things they can figure out how to ML-accelerate as adoption continues, I'm sure. It enables some solutions to hard problems that don't have good deterministic algorithms, and you can run a lot bigger models than you can without acceleration. also fabrice bellard (of course) wrote a neural compressor... https://bellard.org/nncp/ https://bellard.org/nncp/ it's also likely going to be useful in level design and asset creation as well, although of course as we're seeing now with Stable Diffusion there's some interesting legal questions with content creation. optical flow engine (not neural) is another big win for computer vision stuff, too, offloads a few tough tasks to hardware, like object tracking and motion estimation. hooking into that with zoneminder or something would be awesome. I'm also stoked for shader execution reordering too. This is similar to what Intel calls "ray binning", it basically is a best-effort "re-alignment" of the thread (state,etc) to the most similar execution group for memory read alignment/coalescing. I think one or another raytracing implementation (Intel or NVIDIA) they were coalescing based on material. Intel talks about theirs here: https://www.youtube.com/watch?v=SA1yvWs3lHU https://www.youtube.com/watch?v=SA1yvWs3lHU Ada whitepaper: https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvid... I am interested to hear what AMD is doing with RDNA3, supposedly there is a new RDNA3 ISA instruction for matrix acceleration but rumors are it's not a full high-performance matrix unit like CDNA and >=Turing? I don't know why you'd add an instruction without some hardware acceleration though. So maybe less than the other implementations... which is a little disappointing.

wyldfire 4y ago

> I use github copilot every day - its not perfect but good enough to add value to my workflow. Wow I thought this copilot thing was just like a joke. People use it to write software? I'm really curious, now. Can you share examples of the code you're writing with it?

giobox 4y ago
You can use it for just about anything that is text - it will complete configuration files, not just code, and even does a reasonable job a lot of the time writing method docs automatically. Thinking of it as just for code is already a narrow application. Sharing code examples is also of little relevance - just try it yourself and see if it fits your needs. The dollar/value ratio maps for my needs to avoid trips to docs or search engines for syntax confirmation - it may or may not for you. If it gets it wrong sometimes I don't really care, all the time it takes is pressing tab to complete or not to complete if suggestion is garbage. It's just an extension to the tab key's functionality, which is why it's so easy to use every single day when it integrates to most text tools. I also think if this is what we can have today, the future of code completion tools is very exciting.