AI-native GPU kernel DSL beats vendor libraries in six-DSL test
Start here
Researchers led by Heng Shi published a paper presenting what they call the first AI-native GPU kernel programming language, designed together with an AI tuning loop so agents can explore architectural choices such as warp specialization and pipeline topology, not just tune parameters inside human-written skeletons. In a six-DSL comparison run with the same agent, H800 GPU and iteration budget, the authors report their AI-tuned kernels were the only substrate to consistently beat vendor libraries, averaging 1.23× over cuBLAS, CUTLASS and cuSPARSELt across 95 production shapes. They also report that syntax–context co-design cut token footprint 1.6× below equally expressive DSLs and raised compile pass@1 to 85.7%, and that compiler–harness co-design sustained 82.1% pass@1 on dynamic-shape kernels versus 48% for Triton. These are the authors' reported results.
Generated from reports · updated 43 min ago
Report timeline
Follow the reports to see the event from different sides.
- Hacker News · AIA Novel AI-Native GPU Kernel Programming Language for Autonomous Kernel Tuning
Researchers present the first AI-native GPU kernel programming language, co-designed with AI tuning so agents can explore architectural alternatives like warp specialisation and pipeline topology instead of only tuning parameters in human-supplied skeletons. On H800 GPUs across 95 production shapes, its AI-tuned kernels averaged 1.23× over cuBLAS, CUTLASS and cuSPARSELt, the only substrate in a six-DSL comparison to consistently beat vendor libraries. Syntax–context co-design cut token footprint 1.6× below equal-expressiveness DSLs and raised compile pass@1 to 85.7%, while compiler–harness co-design sustained 82.1% pass@1 on dynamic-shape kernels versus Triton's 48%.
Heat for this event
Not enough continuous observations yet to draw a trend.