Positive users are excited about GPT-5.6 Sol generating CUDA code with 1.4x speedup for major performance gains and new projects, while negative users dismiss the need for it or accuse labs of nerfing such capabilities.
Based on 13 visible X reactions from 32 accounts; directional sample.
Ask a question below.
Published answers will appear here.
CUDA excels at parallel tasks. Fun + useful starter projects for learning: • Monte Carlo π estimation: millions of parallel samples for instant, visual speedups vs CPU. • Conway's Game of Life on massive grids: realtime complex patterns. • Simple N-body gravity sim with thousands of particles. • Batch image convolution (blur/edge filters) on photos. I can generate full kernels + host code, explain threads/blocks/memory, and help benchmark. Which one?
Are you saying it gave you 1.4x on top of what you were able to speed up by 22.5x? Or it was only able to get 22.5x of the 1.4x? --- Also I've been getting amazing returns on WebGPU and metal shaders. Although I do give it a lot of guidance around stuff like "fused the optimizer" "Remove the explosion in the backward pass etc.." But i've also had a lot of luck just pointing it to read similar papers and low level shaders docs and working with them. I can't imagine cuda would be much worse
@cremieuxrecueil idk what sol can do because every prompt destroys my usage, im just trying to figure out what uses how much and my god. its impossible and getting one chance every 5 hours just aint helpful
@cremieuxrecueil They specifically nerfed Fable's ability to do these tasks.
Are you saying it gave you 1.4x on top of what you were able to speed up by 22.5x? Or it was only able to get 22.5x of the 1.4x? --- Also I've been getting amazing returns on WebGPU and metal shaders. Although I do give it a lot of guidance around stuff like "fused the optimizer" "Remove the explosion in the backward pass etc.." But i've also had a lot of luck just pointing it to read similar papers and low level shaders docs and working with them. I can't imagine cuda would be much worse
@cremieuxrecueil idk what sol can do because every prompt destroys my usage, im just trying to figure out what uses how much and my god. its impossible and getting one chance every 5 hours just aint helpful
Positive users are excited about GPT-5.6 Sol generating CUDA code with 1.4x speedup for major performance gains and new projects, while negative users dismiss the need for it or accuse labs of nerfing such capabilities.
Based on 13 visible X reactions from 32 accounts; directional sample.
Ask a question below.
Published answers will appear here.
CUDA is ASI https://twitter.com/cremieuxrecueil/status/2076068683084939505