[nikolayasdf123]

> non 100% correctness of kernels

wouldn't model not work properly if kernels are even slightly off?

wasn't kernels a part of training stack for models? am I missing anything?

[arjvik]

I believe their speedup is computed _assuming they can easily fix the correctness bugs in the kernels_.

In practice, with slight differences the model will feel almost lobotomized.

[ymsodev]

The article is referring to GPU compute kernel (https://en.wikipedia.org/wiki/Compute_kernel), not the term kernel used in ML/NN/etc.

[saagarjha]

…aren't they the same thing