Tiny-tpu: A minimal tensor processing unit (TPU), inspired by Google's TPU
github.com267 points by admp 3 days ago
267 points by admp 3 days ago
we are incredibly overdue for the 3D-printer-ification of silicon chips
Don't hold your breath. Chip fabrication is extremely sophisticated. It's basically nanotechnology. Not something that could practically be done at home anytime soon.
I agree, photolithography at home is a long way out.
I think if we want to distribute the means of computational production, a better approach would be to print oligonucleotide instructions (requires a custom inkjet printer), and use a cell free extract (e. coli would work) to synthesize proteins which are programmed to assemble the computational substrate from nanoparticles.
Something along these lines: https://www.pnas.org/doi/10.1073/pnas.1422649112
That being the better approach says volumes about the complexity of lithography.
For anyone reading this, read Chip Wars by Chris Miller. The Lithography sub-story that threads through is amazing and fascinating and the whole book is incredible besides.
For those who think Ai Is a bubble and longshot, look up how crazy of a bet EUV (extreme ultraviolet) lithography was.
I think that before that will be the PCB-ification, where you can order online and be batched in with others on a standardized process, for a reasonable price. This is actually starting to become possible, for example via TinyTapeout you can get tiny chips taped out for a few hundred USD.
There is a company called Nano Dimension working on something like that. Last I heard their 3d printer wasn't at the consumer level yet, but neither were computers in the early days.
That exists, sort of. Just not at an affordable level for home use.
But still worlds apart from current semiconductor fabrication with billions of necessary investments.
https://www.yokogawa.com/industries/semiconductor/minimal-fa...
How did you create such amazing animation with svgs? Cool docs
I was interested in this as well. The site is awesome and beautiful. And, I wished the SVGs were built in a way that (from my limited interaction) is basically just swapping new SVG in for each tick. It's really pretty, but I wished there was a bit more interaction and composability.
Can I get a PCIe card with multiple units and An easy interface? Say 4 TPUs and 8 ram slots? $5,000???
I had approximately the same comment in a thread under a submission about a new CPU <https://news.ycombinator.com/item?id=44201707> and the replies highlighted my understanding of exactly how slow using sticks of RAM versus on-die VRAM. I'd still enjoy trying it, but I have not yet loaded enough of the FPGA ecosystem in my head to know how many hours I am away from watching the bits flow
Those who are expert at this, how many TPUs i need to run lets say
Gemini flash 2.5 and pro 2.5 models for 1 user?
Ah, and a psuedo dupe of https://news.ycombinator.com/item?id=44944592 but I'll be straight that I think the current point spread is accurate because I much prefer the repo link to some random .com domain