Please enjoy this SGI Indigo2 workstation from 1996 running llama2.c by Andrej Karpathy. 1.4 tokens/sec with the 15M TinyStories model! Just a little porting for the big endian IRIX machine, all in an evening’s work. What would ML researchers have thought 28 years ago seeing this?!

By the way, thanks @karpathy for your wonderful repo and tutorials. So thankful for the knowledge you share!

Also, if anyone wants to try this it took way longer just to get this machine going than to port the code lol Many many hours troubleshooting a bad hard drive, installing IRIX 6.5.22 (twice!), configuring, and getting the dev stuff going. Maybe will give this guy a try NeXT

@karpathy this is all extremely tasteful, down to naming this lovely machine "chappie"

@karpathy Thank you!

@karpathy why is this actually fast?

@karpathy Only a 15M parameter model 😉

@karpathy Can you get a 1B LLM running on a computer made in 2000, because that's what Alan Turing predicted.

@karpathy This machine has 1GB RAM. With system overhead, I think it's good for a ~800M as int8. Would require more work to run 4-bit, a lot of bits to bang around but could be done. But, that prediction. That prediction is wild.

@karpathy Pretty mind blowing to consider how AI is written in a decades old language, using decades old math, and utilizing hardware who's fundamental engineering principals are also decades old.

@karpathy I was working in both AI and CGI at that time, using exactly that machine and I would simply have kept looking for the human in the loop. I would never have believed that.