---
type: "post"
id: "forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b"
slug: "forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b"
title: "ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents"
description: "A November 3, 2026 arXiv preprint presents ForgeKernel, a GPU-kernel generation system that carries validated architectural knowledge across optimization runs."
retrieval_nugget: "A November 3, 2026 arXiv preprint presents ForgeKernel, a GPU-kernel generation system that carries validated architectural knowledge across optimization runs. Its authors report competitive results across NVIDIA and Ascend benchmarks, making the work relevant to teams asking whether agentic kernel optimization can improve real workloads rather than merely produce code."
published_at: "2026-11-06"
updated_at: "2026-11-06"
record_date: "2026-11-06"
date_kind: "published_at"
topics: ["compute","inference"]
entities: ["ForgeKernel","OpenBMB","NVIDIA Hopper","NVIDIA Blackwell","Ascend 910C","CANN-Bench","SOL-ExecBench"]
related_patterns: ["agent-economics-moves-to-completed-work"]
novelty_label: "new_fact"
what_changed: "The ForgeKernel preprint adds detailed results: it beats expert references on 25 configurations and reports first-place rankings across CANN-Bench 2.0 and SOL-ExecBench leaderboards."
prior_refs: ["https://newruntime.com/posts/openbmb-open-sources-forgekernel-a-self-evolving-gpu-6db18d/"]
source_url: "https://arxiv.org/abs/2611.04998"
source_urls: ["https://arxiv.org/abs/2611.04998"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b/","markdown":"https://newruntime.com/posts/forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b.md","json":"https://newruntime.com/posts/forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b.json"}
markdown_url: "https://newruntime.com/posts/forgekernel-self-evolving-kernels-catching-and-surpassing-experts-30654b.md"
---

# ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents

## Retrieval answer

A November 3, 2026 arXiv preprint presents ForgeKernel, a GPU-kernel generation system that carries validated architectural knowledge across optimization runs. Its authors report competitive results across NVIDIA and Ascend benchmarks, making the work relevant to teams asking whether agentic kernel optimization can improve real workloads rather than merely produce code.

## Summary
A November 3, 2026 arXiv preprint presents ForgeKernel, a GPU-kernel generation system that carries validated architectural knowledge across optimization runs. Its authors report competitive results across NVIDIA and Ascend benchmarks, making the work relevant to teams asking whether agentic kernel optimization can improve real workloads rather than merely produce code.

## What happened
The paper describes ForgeKernel as a response to two limitations in existing kernel agents: each run begins with the same initial knowledge, and later optimization often stays within an architecture fixed early in the search. ForgeKernel instead analyzes the hardware behavior of a high-performance reference and distills validated architectural decisions into reusable knowledge about hardware mechanisms. The authors say this is intended to avoid simply imitating reference source code.

Its loop uses a validated result as the next reference. Once a generated kernel passes correctness and performance validation, the system “seals” it and uses it to begin another optimization run. That makes it possible to revisit architectural decisions in later cycles, rather than treating the first design as fixed. The supplied primary-source excerpt does not specify access terms, pricing, or a general service offering; it presents the system and results in a research preprint.

For NVIDIA Hopper and Blackwell GPUs, the authors evaluate 27 operators across 29 input configurations. They report that ForgeKernel matches or beats its expert reference on 25 configurations, with a 10% geometric-mean performance gain. The paper also reports results on Ascend 910C: against the per-operator public best for the public case set of all eight level-4 operators in the CANN-Bench Season 1 final leaderboard, ForgeKernel achieves a 25.1% geometric-mean gain. Its full-suite submission ranks first on that benchmark’s A3 leaderboard.

The preprint further states that, as of October 30, 2026, public submissions generated by ForgeKernel ranked first on all three CANN-Bench 2.0 solution leaderboards—Ascend 910C, 950DT, and 950PR—and on the SOL-ExecBench FlashInfer-Bench track. These are claims reported by the paper; the source material supplied here does not include independent reproduction or enough implementation detail to assess the comparisons in full.

## What the archive adds
The archive already covered [OpenBMB open-sources ForgeKernel, a self-evolving GPU kernel optimization agent](https://newruntime.com/posts/openbmb-open-sources-forgekernel-a-self-evolving-gpu-6db18d/), describing the project’s open-source release. This preprint is a new fact on that known story: it adds reported benchmark outcomes and explains the loop of sealing validated kernels for later runs. Earlier coverage of [Mixture-of-Experts kernel engineering on Amazon SageMaker AI](https://newruntime.com/posts/mixture-of-experts-kernel-engineering-on-amazon-sagemaker-abd782/) is adjacent context on GPU-kernel performance, not evidence for ForgeKernel’s claims.

## Trend signal
This supports, with a single research result rather than broad proof, the pattern [Agent economics moves to completed work](https://newruntime.com/patterns/agent-economics-moves-to-completed-work/): the useful measure here is validated kernel performance, not the number of generated kernels or model calls. The reported gains are benchmark-specific; they do not yet establish lower total cost or better performance per verified production task.

## Why now
The timing is tied to public benchmark claims. The paper was submitted on November 3, 2026 and reports leaderboard positions current as of October 30, 2026, including CANN-Bench 2.0 and SOL-ExecBench results. That makes the preprint a snapshot of a live competitive setting rather than a claim about a stable, finished production product.

## Connections to other events
ForgeKernel sits alongside the archive’s [SkillCompiler work on self-evolving agent skills](https://newruntime.com/posts/skillcompiler-self-evolving-agent-skills-compiled-into-executable-c50964/). Both concern preserving useful output for later work, but the artifacts differ: SkillCompiler concerns reusable programs for agent skills, while ForgeKernel uses a validated hardware-specific kernel as the reference for another optimization run. Neither item establishes that self-evolution transfers reliably across domains.

## Why it matters
Kernel generation is a concrete test of whether agents can contribute to performance-critical engineering, where correctness is necessary but insufficient. ForgeKernel’s reported method couples architectural search to correctness and performance validation, then makes a successful artifact the basis for further work. If the approach transfers, teams could spend less expert time repeatedly rediscovering effective hardware-specific designs.

The benchmark figures are promising signals for accelerator software teams, especially those working across different hardware families. But a leaderboard result is not automatically a production win: the relevant question is whether a team’s operators, shapes, compiler stack, and deployment constraints resemble the evaluated cases. The preprint also gives no supplied evidence here about engineering effort, compute costs, or the frequency of regressions. Teams should treat the results as a reason to evaluate the approach on a small representative operator set, not as a guarantee of a 10% or 25.1% gain in their workloads.

## What happens next
- **Independent reproduction will become the decisive test of the reported gains.** (6 months, 60%). Watch for: published reruns on comparable operators and hardware.
- **Teams with repeated kernel-tuning workloads will test retained validated references against fresh-search baselines.** (12 months, 55%). Watch for: reports that include optimization time, compute cost, and regression rates.

## Limits and uncertainty
These are results stated in a preprint, not independently verified findings in the supplied material. The excerpt does not provide enough detail to assess baselines, variance, evaluation costs, or how the selected operators and input configurations represent production workloads. The reported 10% and 25.1% figures come from distinct hardware and benchmark settings and should not be compared as if they were a single test. The source also does not establish access terms, deployment readiness, or the cost of running ForgeKernel. Leaderboard rankings are time-bounded claims; they do not guarantee continued rank or general superiority.

## Skill to build now
Choose one performance-sensitive operator with a stable correctness test and a representative input shape. Record its current latency or throughput, the hardware and compiler environment, and the human time required for each tuning attempt. That creates the comparison ForgeKernel’s claims require: not simply whether generated code is faster, but whether validated optimization improves a real operator enough to justify the evaluation and review loop.

## Reader takeaway
ForgeKernel’s important idea is not just generating a faster kernel once: it preserves validated architectural choices and makes them the starting point for another search. Its authors report sizable benchmark gains, but the practical test is whether the method produces correct, faster kernels on a team’s own operators and hardware at an acceptable cost.

## Open question
Can independent teams reproduce ForgeKernel’s reported gains on representative production operators and measure the total optimization cost?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Compute](https://newruntime.com/topics/compute.md): Explore the compute topic hub.
- topic hub: [Inference](https://newruntime.com/topics/inference.md): Explore the inference topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)