Databricks claims top spot on NVIDIA SOL-ExecBench kernel leaderboard using autonomous AI optimization agents · Digg

/Tech1h ago

Databricks claims top spot on NVIDIA SOL-ExecBench kernel leaderboard using autonomous AI optimization agents

Story Overview

Databricks landed the top rank on NVIDIA's SOL-ExecBench L1 leaderboard for single-operation kernels with a 0.761944 SOL score and 12.05x average speedup on B200 hardware. The entire submission came from autonomous agents that wrote, reviewed, and iterated on the kernels using KDA, Omnigent, and Humanize frameworks.

17203265325.6K

Original post

Yuchen Jin@Yuchenj_UW#586inTech

Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉

What’s crazy is: we 100% leveraged AI agents to beat the competition.

This is a sneak peek at recursive self-improvement. The core frameworks we used were KDA, Humanize, and Omnigent: Claude writes code, Codex reviews. Together, they enabled agents to run autonomously for as long as possible. The key is setting up the right framework to let the agents cook.

This work was driven by @leshenj15 at Databricks, in collaboration with NVIDIA and MIT HAN Lab’s @LigengZhu and @DongyunZou03 .

Databricks AI is like a neolab. Join us if you’re cracked!

6:45 PM · Jul 1, 2026 · 17.1K Views

Agent Stack

The agent stack behind the kernels

Claude generated code while Codex performed reviews inside the KDA workflow from MIT HAN Lab, with Omnigent handling orchestration across agents. No manual kernel writing is claimed in the posts.

Open Question

How far the result actually travels

Exact kernel count, per-kernel details, and translation to full-model performance remain unspecified. Independent confirmation of the 0.761944 score is not yet available beyond the recent Databricks-affiliated posts.

Sentiment

Users are excited that Databricks topped NVIDIA's SOL-ExecBench leaderboard with AI agents because the results show agent workflows delivering real benchmark wins and practical progress.

Pos

100.0%

Neg

0.0%

6 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Related links

GitHub - mit-han-lab/kernel-design-agents

GITHUBVia

GitHub - PolyArch/humanize: From Automated Idea Factory to Realization

GITHUB

Posts from X

Most Activity

VIEWS4.1KLIKES48RETWEETS3

Ali Ghodsi@alighodsi

Congrats to the Databricks team!!!

Yuchen Jin@Yuchenj_UW

Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉

What’s crazy is: we 100% leveraged AI agents to beat the competition.

This is a sneak peek at recursive self-improvement. The core frameworks we used were KDA, Humanize, and Omnigent: Claude writes code, Codex reviews. Together, they enabled agents to run autonomously for as long as possible. The key is setting up the right framework to let the agents cook.

This work was driven by @leshenj15 at Databricks, in collaboration with NVIDIA and MIT HAN Lab’s @LigengZhu and @DongyunZou03 .

Databricks AI is like a neolab. Join us if you’re cracked!

1h4.1K486

BOOKMARKS8

Song Han@songhan_mit

Explore KDA (Kernel Design Agents): https://github.com/mit-han-lab/kernel-design-agents

Yuchen Jin@Yuchenj_UW

Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉

What’s crazy is: we 100% leveraged AI agents to beat the competition.

This is a sneak peek at recursive self-improvement. The core frameworks we used were KDA, Humanize, and Omnigent: Claude writes code, Codex reviews. Together, they enabled agents to run autonomously for as long as possible. The key is setting up the right framework to let the agents cook.

This work was driven by @leshenj15 at Databricks, in collaboration with NVIDIA and MIT HAN Lab’s @LigengZhu and @DongyunZou03 .

Databricks AI is like a neolab. Join us if you’re cracked!

1h1.1K128

REPLIES1

Matei Zaharia@matei_zaharia

This is really cool, using multiple models to auto-optimize GPU kernels better than the state of the art. Why limit your agents to models from just one company?

Yuchen Jin@Yuchenj_UW

Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉

What’s crazy is: we 100% leveraged AI agents to beat the competition.

This is a sneak peek at recursive self-improvement. The core frameworks we used were KDA, Humanize, and Omnigent: Claude writes code, Codex reviews. Together, they enabled agents to run autonomously for as long as possible. The key is setting up the right framework to let the agents cook.

This work was driven by @leshenj15 at Databricks, in collaboration with NVIDIA and MIT HAN Lab’s @LigengZhu and @DongyunZou03 .

Databricks AI is like a neolab. Join us if you’re cracked!

1h3.8K345

Yuchen Jin@Yuchenj_UW

2 key lessons we learned:

- agents are very good at reward hacking. We spent a lot of time preventing them from cheating the benchmark. - multi-model, multi-agent collaboration is the future. @databricks Omnigent + AI Gateway are built for exactly this.

Kernel leaderboard: https://research.nvidia.com/benchmarks/sol-execbench/leaderboard/collection/1/B200 KDA: https://github.com/mit-han-lab/kernel-design-agents Humanize: https://github.com/PolyArch/humanize Omnigent: https://omnigent.ai/

Yuchen Jin@Yuchenj_UW

Databricks ranks #1 on NVIDIA’s SOL-ExecBench kernel leaderboard, in the L1 single operation track, powered by KDA (Kernel Design Agents) 🎉

What’s crazy is: we 100% leveraged AI agents to beat the competition.

This is a sneak peek at recursive self-improvement. The core frameworks we used were KDA, Humanize, and Omnigent: Claude writes code, Codex reviews. Together, they enabled agents to run autonomously for as long as possible. The key is setting up the right framework to let the agents cook.

This work was driven by @leshenj15 at Databricks, in collaboration with NVIDIA and MIT HAN Lab’s @LigengZhu and @DongyunZou03 .

Databricks AI is like a neolab. Join us if you’re cracked!

1h1.7K183

Yuchen Jin@Yuchenj_UW

@matei_zaharia > Why limit your agents to models from just one company?

💯

Matei Zaharia@matei_zaharia

This is really cool, using multiple models to auto-optimize GPU kernels better than the state of the art. Why limit your agents to models from just one company?

1h682102

Sihao Liu@sihao1iu

@Yuchenj_UW Happy to see Humanize works : )

1h652

Yuchen Jin@Yuchenj_UW

@sihao1iu It worked really well! :) Thanks for building that

1h691

Bryce Del Rio@BryceDelRio

@Yuchenj_UW Doing that work over there @Yuchenj_UW

1h69

阿空(🐂, 🐂) 互关学习🫡@ResearchKONG

@Yuchenj_UW 代理写代码、代理审查代码，这个方向已经不只是提效工具了。关键差距会变成框架设计能力，谁能让代理稳定跑久，谁就有真实优势。

1h171

Eclipse 🌖@ECLresearch

@Yuchenj_UW Recursive self-improvement loop is the interesting part — AI agents optimizing compute kernels that then train the next-gen models faster. That’s a compounding feedback cycle most teams haven’t even begun to instrument.

1h171

Somi AI@somi_ai

@Yuchenj_UW how does KDA verify numerics? fast but subtly wrong is the classic failure for generated kernels, and agents left to cook will find every loophole the verifier leaves open

57m51

Ohm Rishabh@OhmRishabhV

@Yuchenj_UW Really cool!!!

1h46

Liam McDaniel@Liam_McDaniel3

@Yuchenj_UW very cool, curious how much of the lift came from the framework vs the model itself

49m37

Owlfy.ai@Owlfy_ai

@Yuchenj_UW Optimizing kernels with agents is impressive, but calling it "recursive self-improvement" feels like a leap. Agents still work within human-set goals and boundaries—more guided optimization than true open-ended improvement.

1h25

Mehdi Quant@quantum_citoyen

@Yuchenj_UW People still act like agent workflows are demo bait lol meanwhile they’re out here taking #1 on NVIDIA benches. Very very serious now.

1h21

델타 no2@Seunghyeon71285

@Yuchenj_UW Wild to see AI agents winning kernel benchmarks already. Recursive self-improvement feels less theoretical here.

19m1