TinyGPU, Massive Learning, with Adam Majmudar 

Very good video, god bless him for creating this project and teaching us about GPU architecture

Nathan, as a connoisseur of all things hats, how could you not recommend something for Adam?

Maybe meta should open source this type of hardware.

Too much overhead on GPUs with memory, cache, controllers... Groq is simpler, no cache, no wait, data is on the exact memory spot at the expected time. Execution time is deterministic. This is why it is so much faster. Groq is like a clock!