GPT-6 Sol writes a CUDA kernel on KernelBench-Mega
Sol [xhigh] wrote a single-launch Kimi-Linear Decode kernel for the RTX PRO 6000 that passed all six checks but ran at 0.96x the optimized PyTorch baseline.
GPT-6 Sol [xhigh] on KernelBench-Mega: It wrote a one-launch Kimi-Linear Decode kernel for RTX PRO 6000. The final result is 0.96x the optimized PyTorch baseline, so this fused version is slightly slower than the baseline. The chart shows its nearby published peers. One cooperative launch runs all four layers: int4 dequantized GEMVs, KDA state updates, MLA attention and routed plus shared MoE. The six-case checker passed. Changing the same input buffer during audit changed the output, which still matched the reference. I haven't used Sol personally yet. I've been abusing Opus 5.5. If you've tried Sol, I'd like to know where your experience lines up with the benchmark.