Edward Z. Yang#811
I've let Fable optimize GPU kernels autonomously using "auto-gpu-kernel" harness, if it joined the NVIDIA's competition today, it would have won 🥇 in 4/5 kernels against humans.
Fable can write Gluon kernels, do warp-specialization, use TMA tcgen05 etc.
(Speedup vs Opus 4.8)
@dogacel0 i wonder how good fable is with cute dsl, it might be able to go more low-level there…
The biggest difference I saw between Opus 4.8 and Fable is Opus can't generate Gluon kernels that are faster than Triton as it fails to handle complexity of warp-specialization or other techniques. As Fable can use a lower-level DSL consistently, it can write faster kernels.
Fable both converges faster and decides to stop earlier unlike Opus. Deciding to stop early might be a safety limit
MoE path had more known optimizations and I've heard people have discovered them by prompting agents continously.
@maharshii Exactly, my harness was designed Opus in mind, I think right now CuTe might be the right abstraction. I will test that when I have time.
Also Fable spawns many sub-agents, utilizes git worktree by itself and overall converges faster. For reference Opus 4.8 kernels took around 12 hours to finish, whereas Fable finished under 8 hours while spending significantly less GPU credits.
Harness + Generated Kernels are available here 👇🏻
https://github.com/Dogacel/auto-gpu-kernel
@dogacel0 yes, considering the fact that one can go to the level of nvvm and ptx if needed in cute DSL
Edward Z. Yang#811I've let Fable optimize GPU kernels autonomously using "auto-gpu-kernel" harness, if it joined the NVIDIA's competition today, it would have won 🥇 in 4/5 kernels against humans.
Fable can write Gluon kernels, do warp-specialization, use TMA tcgen05 etc.
(Speedup vs Opus 4.8)