Chinese AI sweeps NVIDIA's GPU performance leaderboard, with this company claiming 88 first places

By Yunzhong, from Waifeisi Temple

QbitAI | WeChat Official Account: QbitAI

Recently, on SOL-ExecBench [1], a GPU operator performance benchmark launched by NVIDIA Research, Infinigence AI (无问芯穹) achieved impressive results:

It successively topped all four tracks: L1 basic operators, L2 fused operators, Quant quantization operators, and FlashInfer-Bench inference operators.

What is SOL-ExecBench?

SOL-ExecBench is an operator optimization Benchmark launched by NVIDIA Research on March 19, 2026. SOL stands for Speed of Light, meaning the "speed of light"; here it refers to the theoretical performance boundary derived from hardware capabilities.

Unlike traditional benchmarks that mainly compare acceleration effects relative to software baselines,SOL-ExecBench introduces the SOL Score, which uses hardware performance ceilings as the reference, and conducts actual testing on high-performance compute cards.

The leaderboard contains 235 operator optimization problems extracted from 124 real and cutting-edge AI models, covering language, vision, speech, video, multimodal and other domains, divided into four tracks:

• L1 basic operators:Testing efficient implementation of basic computational modules.
• L2 fused operators:Testing fused execution and coordinated optimization of multiple computational steps, reducing memory access, data movement, and execution overhead.
• Quant quantization operators:Covering low-precision computing technologies such as FP8 and NVFP4, closely related to the performance and cost of large model inference.
• FlashInfer-Bench inference operators:Focusing on key computations in large model online inference, such as attention mechanisms and mixture-of-experts networks.

A notable feature of SOL-ExecBench is thatit extracts workloads directly from the computation graphs of real cutting-edge models, and focuses on testing the performance limits of these computations on next-generation GPUs。

Note: This is a continuously competitive global leaderboard with frequent dynamic updates. As teams keep submitting new optimization solutions, the rankings change accordingly, and the competition is fierce.

As of October 26, 2026, Infinigence AI still ranks first globally in total scores on both the L1 and Quant tracks, and second globally in total scores on the L2 and FlashInfer-Bench tracks; among all 235 GPU operator optimization problems,with 88 first places on individual problems, it tops all participating teams, more than 2 times the number of first places of the second-place team.

Especially on the L2 fused operator track, which tests complex computational optimization capabilities, Infinigence AI won, out of the 82 problems,39 first placesout of 82 problems, nearly half.

This competition, which brings together the world's top AI infrastructure and agent development forces, is becoming an important stage for testing next-generation AI compute optimization capabilities.

From GPU Maxxing to Token Maxxing, a global competition for converting AI productivity

The realization of AI productivity value lies behind a complete conversion chain of "investment - electricity/cards - tokens - productivity - value".

Driven by the early Scaling Law, investing more computing resources and maximizing GPU efficiency were important means to push breakthroughs in model performance and increase token production.

And with the development of Agents, Tokens have become the input for task execution, rather than the final deliverable,Models can break through problems that were originally hard to solve through longer reasoning and more rounds of attempts and verification, but simply increasing Token input does not necessarily bring improvements in task performance.

If similar attempts are always repeated and direction cannot be corrected from execution feedback, no matter how many Tokens are consumed they may be spent spinning in place.

The industry's core focus is further shifting from GPU Maxxing to Token Maxxing:It is not only about improving the conversion efficiency from "electricity" to "Tokens", but also about turning more Tokens into real model capability and task outcomes, achieving efficient conversion from "Tokens" to "productivity".

SOL-ExecBench provides concrete scenarios for testing this capability.It extracts computational workloads from real frontier models, covering basic operators, complex fusion, low-precision computation, and inference optimization, requiring submitted code to satisfy both correctness and performance requirements. Here, Agents must go through code generation, compilation and execution, performance analysis, and repeated tuning, with their work ultimately measured by execution results on real hardware.

This benchmark attractedparticipation from technology teams including Databricks, a global data and AI platform company; SF Tensor, an AI Infra startup incubated by Y Combinator; Cursor, the well-known AI coding tool team; as well as Recursive, doubleAI, Geometric, and other teams active in automated operator optimization.。

According to data published by SF Tensor on September 25, 2026, at that time already583 teams had cumulatively submitted more than 31,000 evaluations, at an average of about one public submission every 8 minutes, continuously driving the leaderboard refresh.。

Many participating teams used Agents to automatically generate, test, and optimize low-level computational code, making this competition a test both of GPU Kernel optimization skill and of how to organize model capability into a system that can continuously complete complex engineering tasks.

In this fierce global contest, Infinigence AI successively achieved first place worldwide in the four major tracks: L1, L2, Quant, and FlashInfer-Bench.

Thisnot only reflects the team's technical accumulation in operator optimization and software-hardware co-design, but also tests the team's ability to convert Token input into verified computational performance gains through Agentic Infra, thereby exploring effective paths for AI to autonomously complete complex tasks.。

Using AI productivity to feed back into Infra evolution, and technical practice to help RSI engineering land in practice

From resource investment to Token production, and then to the delivery of actual productivity and the realization of value, Infinigence's Agentic Infra has always been committed to pushing this pipeline from one-way advancement toward closed-loop evolution.

The team hopesto let Agents not only consume Tokens to complete tasks, but also feed the productivity they create back into the infrastructure itself, continuously improving Token production and exploring the path toward realizing RSI for AI infrastructure.

Operator optimization tasks provide a concrete entry point for exploring this process:The execution efficiency of GPU Kernels directly affects how hardware performance is unleashed; once validated and applied to training and inference systems, the efficient implementations found by the Agent are expected to further reduce computational overhead, extending the benefits of operator optimization to the Token production stage.

Along this direction, the first thing to solve is how to enable the Agent to continuously produce reliable optimization results. Whether code executes correctly, where the performance bottleneck lies, and whether a single modification brings real gains all need to be judged through actual execution and become the basis for the next round of exploration.

To this end, Infinigence built an operator optimization Agent around SOL-ExecBench, organizing code generation, execution validation, performance analysis, and experience accumulation into a continuous workflow.

Below, from three aspects—containerized isolated execution, Skills perception with shared Memory, and optimization strategies for different tasks—we introduce how Infinigence concretely supports the Agent in achieving continuous improvement under real hardware feedback.

Containerized Isolation Architecture

In terms of execution architecture, the system adopts containerized isolation, decoupling the Agent workspace, the evaluation zone, and the debugging zone.In a workspace with network access allowed but no directly allocated GPU, the Agent loads Skills, shared Memory, and optimization strategies via external mounts, completes code writing, and submits computation requests.

A host-side resource scheduler centrally manages the GPUs, queuing requests and allocating them on demand; evaluation and debugging are executed respectively in an image aligned with the official evaluation environment and in an image with integrated performance profiling tools. Both types of execution containers have network access disabled, and results are returned to the Agent, forming an iterative closed loop of "code generation—evaluation validation—performance analysis—strategy adjustment."

This isolation design can serve toprevent the Agent from bypassing evaluation constraints during the automatic optimization process.When the Agent takes performance metrics as its optimization objective, its exploration may deviate from operator implementation itself and instead attempt to modify evaluation scripts, interfere with timing logic, access reference results, or exploit environment loopholes to obtain scores that do not come from genuine improvements in computational efficiency.Container isolation, permission control, and a controlled submission interface together constitute the execution boundary, allowing the Agent to continuously optimize candidate operators while limiting its interference with the evaluation process itself.

Modular Skills and Shared Memory Mechanism

In building Agent capabilities, the system not only taps into the technical accumulation of the open-source community through modular Skills, but also uses a dedicated framework-aware Skill to help the Agent understand and use the relevant frameworks, while leveraging shared Memory to continuously accumulate and reuse optimization experience.

• Knowledge base and performance debugging Skills: from the open-source community, providing the Agent with capability support such as GPU programming knowledge, performance analysis methods, and optimization tools.
• Runtime-framework-oriented interaction Skill: lets the Agent in the workspace clearly understand its execution environment, the capabilities it can invoke, and the boundaries of its operations, initiating debugging, submitting evaluations, querying task status, and reading results through a unified CLI entry point.
• Shared Memory: used to record strategies, applicable conditions, and practical experience verified effective during actual optimization, enabling different tasks to reuse existing exploration results.

Together, these mechanisms drive the Agent to invoke professional capabilities, correctly use the runtime framework, and accumulate experience and reduce repeated exploration through continuous iteration.

Fine-Grained Optimization Strategies Targeting Task Characteristics

In terms of optimization strategy, the system organizes layered exploration around workload characteristics, performance bottlenecks, and time budgets.

The system guides the Agent to choose appropriate programming abstractions and DSLs (domain-specific languages) based on the optimization time budget, weighing development efficiency against performance potential among DSLs of different granularity such as Triton, TileLang, and CuTe DSL, and continuously mining the hardware performance space by generating lower-level languages (such as CUDA, PTX) layer by layer.

For different computing scenarios, optimization strategies are further refined:

• Compute-intensive tasks: Focus on evaluating Tensor Core utilization, tiling schemes, and data reuse, exploring asynchronous data movement and pipeline overlap, while balancing register and shared memory usage against effective parallelism.
• Memory-intensive tasks: Prioritize reducing redundant reads/writes and intermediate results landing on memory, optimize merged memory access, vectorized access, and cache reuse, and reduce the amount of data movement required per unit of effective computation.
• Multi-operator computation chains: Evaluate the benefits of operator fusion for memory access and launch overhead, while also paying attention to register pressure, synchronization complexity, and changes in parallelism after fusion. For applicable dependency chains, explore Programmatic Dependent Launch (PDL), allowing work in successor Kernels that does not depend on prior results to execute early, with necessary synchronization completed before reading dependent data; for tasks suited to persistent execution, explore Persistent Kernels, which keep CTAs continuously resident and loop over multiple work units, amortizing thread block scheduling and initialization overhead, and further improve execution efficiency through pipelined parallelism.
• Small-Shape, low-latency tasks: Focus on controlling the number of thread blocks, invalid threads, boundary handling, and instruction and synchronization overhead, avoiding excessive tiling or complex pipelines that introduce excessive fixed costs that cannot be amortized by effective computation.

By combining open capability reuse, layered performance optimization, and real hardware feedback, Infinigence hopes to transform GPU operator tuning, which is highly dependent on expert experience, into an optimization process that Agents can autonomously explore, continuously learn from, and execute at scale.

Expanding the boundaries of intelligent evolution with continuously innovative Agentic Infra

The achievements of SOL-ExecBench are not an isolated technological breakthrough, but a natural extension of Infinigence's long-term deep commitment to Agenetic Infra technological innovation.

From connecting "M models" with "N chips", to the Agentic Infra strategic layout of "front-end store, back-end factory, and one center" centered on "diverse heterogeneity, software-hardware co-design, and autonomous AI", Infinigence has always focused its technical exploration and practice on the core links of the AI productivity chain.

This exploration and these achievements on SOL-ExecBench are not only a phased manifestation of its long-term technical路线 in the field of low-level computation optimization. More importantly, they demonstratethe feasibility of the RSI path of further handing this efficiency optimization capability to AI itself, initially realizing an evolutionary closed loop from Token production to productivity creation, and then from productivity feeding back into Infra.

Next, Wuwen Xinqiong hopes to extend this exploration to the task execution framework of the Agent itself. By modifying and validating task planning, tool invocation, memory management, and feedback mechanisms, the Agent can gradually improve its own problem-solving methods as it completes tasks.

When optimization results not only improve computational efficiency but also improve the mechanism that produces the optimization solutions, continuous exploration has the opportunity to be transformed into continuous capability growth.

Facing the future evolution of RSI and AGI, while being full of confidence, Wuwen Xinqiong's technical team is also clearly aware that single-point technical breakthroughs still have a long way to go before reaching true system-level capability, and turning local optimizations into real productivity likewise requires more technical exploration and collaboration.

From making AI use computing power more efficiently, to letting AI participate in the optimization of the computing power system itself, Wuwen Xinqiong is continuously pushing infrastructure from supporting AI innovation toward co-evolving with AI, exploring new paths for the sustained delivery of effective AI productivity.

Acknowledgments:
Humanize community (https://github.com/humanfia/humanize)

References:
[1] SOL-ExecBench: A speed-of-light benchmark that measures real-world GPU kernels against hardware limits. Published in 2026. Available at https://arxiv.org/abs/2603.19173
[2] KernelBench: Can LLMs Write Efficient GPU Kernels? Presented at ICML 2025. Available at https://arxiv.org/abs/2502.10517
[3] MultiKernelBench: A benchmark for kernel generation spanning multiple platforms. Published in 2025. Available at https://arxiv.org/abs/2507.17773
[4] CANN Bench: Evaluating agent-generated kernels against real NPU and algorithmic limits. Published in 2026. Available at https://arxiv.org/abs/2607.20518

*This article is republished by QbitAI with authorization; the views are those of the original author alone.