Hacker News
Show HN: Andrej Karpathy's microgpt.py to C99 microgpt.c – 4,600x faster
Andrej Karpathy's microgpt.py was reimplemented in C99 as microgpt.c, achieving 4,600x faster training. The C version uses SIMD auto-vectorisation and INT8 quantisation, reducing weight storage by 8x. This results in significant efficiency gains, with training 2,300x faster on Windows and 4,600x faster on a MacBook Pro M2 Max.