Home made GPU escalated quickly [video]
youtube.com110 points by erichocean 3 days ago
110 points by erichocean 3 days ago
I love the absolute insanity of this design. And the Thinking Machines vibe.
Computers need more blinking lights.
Very cool project.
A great advice one gets at around minute 9 is to place footprints for anything you consider remotely possible or that you'd like to test. You can always leave them unpopulated and the tradeoff between area lost and time lost is usually worth the area, especially in the first iterations of a pcb.
So much in there: so much hand-soldering of SMD, the way he made an SMD resistor bridge to bodge his MOSI/MISO mixup, using the Bambu 3D printer as a test harness (with pogo-pin attachment) to test his "blades"…
(I thought he was going to end up with R2-D2; the way the design was going…)
The automated programming fixture/conversion was pretty sweet, im not gonna lie about that.
This is great. It somewhat reminded me of Steve Ciarcia's build of a Mandelbrot-generating supercomputer from around 1990. That was also made from microcontrollers (Intel 8052 in that case).
Do you have a link? As far as I can tell you're referring to a BYTE magazine article, but I can only find vague references to it, not anything specific enough that would allow me to locate it.
It was in Byte circa 1989-1991 & I've been trying to search for it with no luck so far.
FOUND IT! https://archive.org/details/1982-10-byte-magazine-october-1-...
The determination to pull through a project of this scale is mind blowing and the joy is contagious.
For me it's the skill. So many things you have to be good at to build this. Really impressive and slightly jealous.
Yeah I don’t think I could ever commit to a project like this and of this scale.
Might be different if it was something truly useful or novel vs a nerd snipe.
Very impressive that he pulled it off in a relatively short amount of time.
Im curious what FLOPS and per CPU bandwidth this has. It might be okay at running compute intensive shaders!
0 FLOPS those MCUs only support IMBC instructions. No hardware floating point, at least it has integer multiplication/division. My estimate is maybe 1-8Gflop total using software float. If you avoid float and design around fixed point, might do some interesting stuff.
IIRC, the Connection Machine CM-1 also was not designed for floating point ops, but Richard Feynman convinced Danny Hillis that it could compete with supercomputers in the scientific market.
@3:10 “you need to consider 99.9% of the power is converted to heat” uhm that would be quite an efficient heater you designed there!
One could argue that it is 100% as every Watt that enters through the socket (not counting reactive energy) is going to become heat at some point in the very near future, including the air moved by fans, the photons emitted by the screen, etc etc.
Not really? The best heaters are heat pumps which can under typical operation reach 300-500% efficiency (COP). Technically they aren't converting the electricity into heat, but for arguments sake here about heaters we only care about electricity consumed to heat generated.
> Technically they aren't converting the electricity into heat
In fact they are converting 100% of it, they are just also collecting and moving additional heat in the process.
Isn't that lower than a purely resistive heater?
Only if the resistive heater in question is DC or AC with reactive power compensation
A purely resistive heater, by definition, has no reactive component so there's no reactive power considerer.
The 0.1% mentioned might be the light that the project produces.
Assuming you're not in an endless vacuum, the light will also become heat.
I think it's fair to separate it from the direct heat generation because if you look at second order effects like that, pretty much all of the energy ends up being radiated away as infrared in space.
It's hardly a second order effect if you're talking about heating a space and the light is absorbed by the walls of that space.
It's quite literally a second order effect. The first order effect is light generation, and a consequence of that effect is heat generation via light absorption.
Every computer is just space heater with side effect of computations.
The same way every diesel engine is just oil stove with side effect of rotary motion. If the engine was in the back of the car you could totally put a pot on it and braise something.
A diesel engine is a very inefficient stove though, as only half (more or less, depending on the exact engine) of the energy is converted to heat, the leftover being mechanical energy.
Modern yes. But olden times atmospheric diesels without turbo were really efficient space heaters.
Good ol' days.
The Volkswagen 1.6D is probably the most reliable engine ever made. Only two failure modes - overheat or shitty fuel for the span of decades to break the fuel pump.
I’m curious about what you’re thinking of, but for both electric heaters and computers, essentially 100% of the input energy is converted to heat one way or another.
What you may be thinking of is efficiency when the output is intended to be something other than heat. In those cases, efficiency is lost because a significant proportion of the input energy is converted to heat.
But if heat output is what you’re interested in, I’m happy to report that 100% is a perfectly achievable, in fact hard to avoid, number!