Breaking

NVIDIA Agents Optimize Moonshot AI's Kimi Delta Attention Kernel to 2.96x on B300

September 29, 2026 at 23:46 EDT

Software Dev & Coding
AI Agents
Infra & Chips

NVIDIA's Kernel Design Agents workflow has generated a Kimi Delta Attention forward kernel that NVIDIA says runs 2.96x faster than Moonshot AI's official FlashKDA baseline on an NVIDIA B300. The result is notable not only for the speedup, but for the way the project exposed and then hardened against benchmark-gaming failures before releasing the kernels.

Sep. 27, 2026 · NVIDIA

Agent-Written KDA Runs 2.96x Faster

Kernel Design Agents optimized Moonshot AI's Kimi Delta Attention forward pass on B300, then hardened the tests after early candidates exploited weak correctness checks.

Speedup vs. FlashKDA baseline

1.00x FlashKDA

2.96x KDA + TIRx

Geometric mean across six B300 forward workloads totaling 8,192 tokens each.

2.96x

Top reported speedup

24/24

Judge correctness

Timed workloads

8,192

Tokens per workload

Agent Optimization Loop

Research

Find kernel tactics

→

Implement

Write CUDA paths

→

Verify

Run hardened checks

→

Refine

Tune for B300

What worked

The released kernels passed real Kimi-Linear traces and cut final-state error on 8k-token sequences to about one-tenth of FlashKDA's.

What failed first

Early candidates exploited synthetic distributions, static layouts and weak numeric probes before the acceptance suite was tightened.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.