The Databricks team said AI the system ranked NVIDIA SOL first in all four tracks of -ExecBench
BIBIBI
AT A GLANCE
The Databricks team said its AI ranked first in all four tracks of 235 kernel tests,GPU the kernel tests.
Article
For Databricks AI Sweeping NVIDIA's Kernel leaderboard: 235 first place in all four tracks across tests,Token costing approximately 70000 USD Beating AI Flash News, the Databricks team said they put GPT-6 Astra and Opus 5 into an automated optimization loop, allowing the models to continuously rewrite, test, and improve GPU kernels, ultimately ranking first in all four tracks of NVIDIA SOL-ExecBench. This benchmark contains 235 kernel tasks, covering basic operators, complex fused operators, low-precision computing, and real-world large-model inference. The team said the models GPU throughout the process Token cost approximately 70000 USD.
This system is built on KDA and Humanize, enabling AI to write kernels, run correctness and performance tests, and continuously modify them based on the results.KDA The team recently applied a similar method to Kimi Delta Attention, achieving on NVIDIA B300 2.96 times the speed of Moonshot AI's official FlashKDA.
The two achievements are different applications of the same Agent kernel optimization system. Kimi demonstrates that a highly challenging attention kernel can be rewritten by AI; Databricks' work extends this method to hundreds of different kernels.
Original link https://m.theblockbeats.info/flash/369536
Key points
01
The Databricks team said it put GPT-6 Astra and Opus 5 into an automated optimization loop.
02
The system enables the models to continuously rewrite, test, and improve GPU kernel。
The Databricks team said the models in this process Token cost approximately 70000 USD。
05
This system is built on KDA and Humanize, and performs correctness and performance testing.
06
KDA The team said a similar method on NVIDIA B300 enabled Kimi Delta Attention to reach 2.96 times the speed of Moonshot AI's official FlashKDA.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This is a system that enables AI to automatically write and optimize GPU kernels: the model first generates code, then runs correctness and performance tests, and continues modifying it based on the results. The article says this method was used for Kimi Delta Attention and expanded to multiple GPU kernel tasks tested by the Databricks team.
Why it matters to readers
If the reported results can be reproduced,AI automated optimization GPU of kernels could improve hardware utilization and demonstrate the ability to handle large numbers of different kernel tasks.
GPU A kernel is low-level computational code running on GPU. Automated optimization means AI can participate in writing, testing, and improving this type of code, but the results in the article remain primarily the team's own claims.
Risks and unknowns
First place in four tracks, approximately 70000 USD Token in costs, and 2.96 times the speed all come from statements by the relevant team.
The article does not provide complete test configurations, code, or independent replication results.
The article does not state whether this method has entered actual products or customer deployments.
Related Developments
Loading event timeline…
Related concepts
GPU Kernel
This term is not in the glossary yet. Browse related concepts in the glossary.
