Pinned post
Joe Fioti
Verified account This account is verified. Learn more@joefioti
521 Following 3.7K Followers 2.1K Posts
Latest posts 5
For some reason (probably my GPU MODE talk) people think Luminal is a megakernel™️ company.
We are not! They are one (of many) possibilities our compiler searches over, and are often not the right solution for a given workload. Sometimes they are, sometimes they aren’t, and it’s often not knowable beforehand. Search wins here!
We are not! They are one (of many) possibilities our compiler searches over, and are often not the right solution for a given workload. Sometimes they are, sometimes they aren’t, and it’s often not knowable beforehand. Search wins here!
Last Year We Built Megakernels. This Year We Deleted Them.View quoted post on X
One year ago, we helped introduce some of the first AI megakernels to the world. Since then, we’ve enjoyed seeing their wide adoption from friends and researchers. However, while whole model megakernels are a helpful abstraction, we’ve found that they overcomplicate and are not always an ideal optimization target.
Below, we discuss our journey to understand and push past these limitations. We start with a deep dive into two contrasting paradigms of ML optimization – Torch Compile and Megakernels
Big agree
Fireworks has told investors that chips and data centers are a commodity and that its real moat is software that runs models quickly.View quoted post on X
Read AI Agenda: thein.fo/4iJEGME
I’ll be on TBPN this afternoon to talk about heterogeneous inference
Happy Thursday. On today's show:View quoted post on X
- @dylan522p (SemiAnalysis)
- @EricNewcomer (Newcomer)
- @tantumscollins (Inherent Labs)
- @joefioti (Luminal)
See you on the stream.