tinygrad's Llama 8B training time reportedly drops to 102.5 minutes on AMD MI350X
tinygrad compares its latest result with AMD's 108.9-minute time and says it also has multi-machine training working.
TLDR
On September 14, tinygrad reported a 102.5-minute Llama 8B training time on AMD MI350X, improving on the 108.5 minutes it reported on September 9. In the earlier announcement, it put the top MLPerf time at 108.9 minutes and said its own machine was slowing down from heat, with fans but no air conditioning. The update also says multi-machine training is working and argues that further speed gains are possible.
Combined views
23.4K
1 Source, first seen 16d ago