if it can't run on a raspberry pi zero 2W, even very slowly, it's bloated
sure
>>102419862>surehttps://github.com/vitoplantamura/OnnxStreamyou can run Stable Diffusion XL 1.0 on the rpi zero 2W, with 298MB of RAM, there's no excuse
>>102419644How do you power it?
>>102419980Does this include training or are these prebaked models?
>>102419989https://datasheets.raspberrypi.com/rpizero2/raspberry-pi-zero-2-w-product-brief.pdf5V DC 2.5A micro USB
>>102419980> Generating a 10-steps image takes about 11 hours on my RPI Zero 2.This shit isn't running. It's not even walking. It's crawling on a top broken glass.
>>102419994>traininghttps://en.wikipedia.org/wiki/Stable_Diffusion>The model was trained using 256 Nvidia A100 GPUs on Amazon Web Services for a total of 150,000 GPU-hours, at a cost of $600,000you can't train image models on anything less than a datacenter
>>102420017>even very slowlywhat do you propose as an alternative benchmark for being bloat free? programs nowadays crash if you don't have more than 8GB of ram just to run a fancy electron app, the fact the SD runs even if at snail pace, is impressive at all. for a $15 computer with the average developer technical competence, it shouldn't even be capable of finishing but it somehow does because the software was written well
>>102420019I see thanks for the quick rundown OP
>>1024200172W though. That's 2 free images a day.
>>102419644Based, but avoiding bloat is unrealistic for most applications.
>>102419644i wrote an opengles 2.0 renderer for 3d models, the pi zero 1 was pushing 10,000 tris no problemmodern software is cancer. btw raylib didn't run on it at all. i just wrote it in pure c and opengles and x11. 1 line of code to compile.
>>102420083you stupid.2W/24H = 4W/12H = 8W/6H = 16W/3H = 32W/90min = 64W/45min = 128W/22,5min = 256W/11,25minMy pc with rtx3060 pc uses less than 256W and is able to generate an equivalent image in around 2 seconds. It also allows me to use better models.
>>102420792both of you are midwits. programming for a raspi zero requires an entirely different brain than any of the shit out there on githubgithub code spend all of its time chasing pointers, waiting for data to transfer, and evicting cache lines needlessly>t. what is a pre-allocated array
>>102420792imbecile
>>102420842explain what you mean then
>>102420817Didn't read because rude.
>>102420792Are you retarded on purpose?
>rpi zeroFor those that are into hardware, how does ARM fair in the energy efficiency to performance ration areas? Maybe also factoring in ease of programming for / ISA.ARM and RISC-V came out more recently than the last time I looked into these kinds of things.If you didn't care about legacy / ideal situation, what kind of platform would you prefer to have for the kinds of applications you'd want to utilize it for?
>>102420792What is this algorithm called?
Modern 3d game engines have a lot of overhead, lots of stuff going on behind the scenes. You can open up a brand new project in any of the most popular 3d game engines and just make a game about looking at a static primitive object in an empty scenes and you'll likely only get hundreds of frames per second. Realistically, you should be getting tens of thousands of frames per second rendering what is essentially nothing at all, but it won't turn out that way. I don't think an Unreal 5 engine "empty 3d game where you just look at a static cube" would run well on a pi zero
>>102421204the main reason to use arm is if you care about cost or thermals at all. performance , even with careful c programming that most "engineers" today are scared to do, will always be quite bad. but the only applications that need good performance on new features are top end ones like the latest games, the latest ML models, and servers under high load.to illustrate how bad it is, hello world in java on a pi zero takes minutes to run. 99% of programmers these days are afraid to write even a single leetcode problem in c
>>102421297>... but the only applications that need good performance ...Neat. I've been considering making a Plan 9 cluster where these would basically be thin clients. Low thermals and low intensity processing client side, all heavy lifting done on other nodes.>Java on piI don't have experience with Java but I imagine the JVM used matters here. Has there been work done on making an optimal implementation for embedded arm devices like that? I'd imagine less than has been done for older platforms.I do know SUN used it on their smart cards.https://en.wikipedia.org/wiki/Sun_Raywhich Oracle still produces today.Not sure what these things were capable of in terms of compute and memory.
>>102421237>bideo gaymesBloat
>>102420019256 GPUs is a rack, not a data center.
>>102420702source code or it didn't happen
>>102420702>btw raylib didn't run on it at all.Did you use the correct backend?
>>102419644All the cycles you waste in your head deciding what software other people are allowed to need is bloat. Your brain, and your ego, are extremely bloated and putrescent, like a corpse
>>102421216>>102421186>>102420842is /g/ really this retarded? or is that retard samefagging?running 100watts for one hour costs the same as running 10 watts for 10 hours.
>>102419980runs way faster on an fpga but lets be real, OP, youre a faggot that doesn't even do anything with their hardware
>>102422222a shitty solar panel can output > 10W for most of the day. it cannot output 100W for most of the day however.
>>102421816imagine not knowing how to create a window in x11, a glsl shader for sampling a texture, an obj parser that uses a simple hashing function, and basic projection and rotation matrix code available in thousands of repos, and some calls to submit bind and draw calls and a present to a surfaceits literally 500 lines
>>102422183if a self proclaimed "easy to use library" cant successfully compile and run when my many thousands of lines of code opengles and vulkan code can, then your "library" is dogshit and i will happily continue living my life unafraid of big bad scary uwu monsters like reading a gpu api spec and writing one line of gcc to compile rather than relying on "totally necessary build systems" like make or cmake or ninja or meson or vscode or
>>102422231>it cannot output 100W for most of the day however.generating 2 images on a proper setup takes less than 10s, closer to 5(at 200w).using a pi for this is insanely retarded.
>>102422222lol, retard
>>102422291bruh. i just read the op again
>>102421297Nah, M1 is arm and it's performance is very good compared to the i7-9700. Perhaps a little faster. This despite much lower clocks and fewer proper cores and 1/10 the power usage.
>>102419644I actually have a pi400 I got for christmas a few years back that I use as a "worse-case scenario" kinda device to test our (company) software on2/3rds of the team I help fucking babysit absolutely despise me because it means they actually have to try and can't just rawdog bloated chatgpt AI slop but everyone else keeps making jokes about how I really shouldn't leave and my boss pulled me aside and got really nervous once because I wore an old microsoft hoodie I found at a yard sale and he was extremely concerned it meant I was gonna leave.
>>102422363M1 arm is not raspiM1 arm is not epycM1 arm is not 5.0ghz singlethread perfM1 arm is extremely efficient at doing only what it was designed for— 100 tabs of YouTube and google docs
>>102419644If it can't run on an iPod nano 1g, even very slowly, its bloated
>>102422276Skill issue You didnt change the opengl backend defualt is opengl 3.3 while the pi zero w has opengl es 1.0 and 2.0 support
>>102422581Kill yourself, you can’t even write a 3d application without someone giving you a scene graphBy the way, detecting what devices and APIs are available on a host system is literally trivial and should not require someone to “CONFIGURE A BACKEND” to use it. Kill yourself. Die in a fire AIDS monkey faggot
>>102419644Was watching a thing the other day where a guy stripped away all of the bloat on a mission critical component that needed to be a performant and reliable as possible, because of latency issues.It took two weeks for the other devs to cry and make management force him to add all the B.S back in instead of using the replacement calls that he implemented.
>>102420702raylib is a toy project. Its not meant for serious stuff but for education sokol is better or sdl3 now.
>>102422232so lets be clear here, you are lyingyou sir are a liar