MLCommons' AI training tests show that the more chips you have, the more critical the network that's between them.

Using a clever solution, researchers find GPT-style models have a fixed memorization capacity of approximately 3.6 bits per parameter.