Question on reported numbers

[LeoXinhaoLee]

Hi, thank you for releasing code for this cool project. I have the following questions for numbers in the paper

1. How is Fig. 9 produced?

• What’s the variance over? Caption says “evaluation environment”, what does that mean?
• The mean speedup in Fig. 9 seems to be different from Table 4, is this expected?

2. The repo releases some highlighted kernels and a csv recording their performance:

• Are these kernels the best kernels in Fig 4? (e.g., mnist_linear_forward , layernorm csv obviously doesn’t match Fig 4, others seems reasonably close)

3. What’s the exact and complete environment besides Table 6? Is there a specific cloud platform and docker image to reproduce numbers in the paper?

Thank you very much for your time and help!