His kernel-tuning agent loop fixes the numbers and has not yet hit the speed target
His automated agent loop optimizes kernels and compares CPU oracle against CUDA outputs layer by layer, solving the numerical problems but not yet matching the target speed.
This only partly worked. Read what happened before you try it.
How it works
- An automated agent loop optimizes kernels
- Compares the reference implementation's outputs against the GPU outputs layer by layer
- Solves the numerical problems
- Has not yet matched the target speed
Why it’s worth knowing
Right but slow is a real stopping point, and he reports it as one.
Where it’s from
Summarized from a public post on news.ycombinator.com. See the original ↗