SRIBD News

Editor’s Note

Operators—the smallest computational units that support AI model inference and training—serve as the core bridge between AI algorithms and compute chips. Their efficiency and compatibility directly affect model runtime speed and deployment difficulty. For a long time, operator development has faced challenges such as high barriers and long cycles, making it one of the key bottlenecks constraining the development of China’s domestic compute ecosystem.

Zhizi Xinyuan (智子芯元), a company incubated by the Shenzhen Research Institute of Big Data (SRIBD), has independently developed an intelligent tool called KernelCAT, exploring a new technical path for operator development and performance optimization.

Recently, Zhi-Quan (Tom) Luo, Academician and Director of SRIBD, sat down for an in-depth interview with Greater Bay Area Voice (大湾区之声) of China Media Group (CMG). In the conversation, Luo systematically explained KernelCAT’s technical philosophy, application prospects, and his deeper reflections on building China’s domestic AI compute ecosystem. He stressed that for the domestic AI ecosystem to move from “following” to “leading,” the key lies in short-term tool breakthroughs that make users willing to adopt, medium-term ecosystem refinement that makes them use it well, and long-term talent cultivation plus open collaboration to form a virtuous cycle.

We reproduce the interview below.

(Image: CMG Greater Bay Area Voice 《科创梦工厂》program card / studio still)

About Zhi-Quan Luo

Zhi-Quan Luo is Director of the Shenzhen Research Institute of Big Data, Vice President of the Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen), Executive Dean of Shenzhen Hetao Academy, Foreign Member of the Chinese Academy of Engineering, Fellow of the Royal Society of Canada, IEEE Fellow, and IAM Fellow. His research focuses on optimization methods for big data analytics, as well as complexity and computation in signal processing and digital communications. He proposed data-driven statistical simulation techniques for real-world networks, built mathematical models and algorithm frameworks for large-scale 4G/5G heterogeneous network parameter optimization, and broke through algorithmic bottlenecks in solving ultra-large-scale mixed-integer optimization models. He is a recipient of the IEEE Signal Processing Society Best Paper Award (2004, 2009, 2011, 2015), the IMA Paul Y. Tseng Memorial Award (2018), the CSIAM Wang Xuan Applied Mathematics Award (2022), the Shenzhen Science and Technology Progress Award (First Class, 2023), and the ICCM Hua Luogeng Prize (2025).

Click “Read More” at the end to listen to the full Greater Bay Area Voice 《科创梦工厂》program.

▐ Why KernelCAT is a “Key Booster” for China’s Domestic AI Ecosystem

GBA Voice:​ Could you introduce KernelCAT’s core positioning and how it works?

Luo:​ KernelCAT’s kernel positioning can be understood as a “compute acceleration expert” inside the AI compute ecosystem. It targets the three most painful bottom-layer problems when AI models go into production:

1. Large language models fail to run after being migrated to domestic chips;

2. They run, but slowly;

3. The development cycle for operators—the minimal compute units behind AI inference and training—is too long.

KernelCAT’s goal is to turn engineering work that used to take experts weeks or months of trial-and-error into an intelligent service: you describe your need in natural language, and mathematical tools plus AI tools automatically deliver the result. Technically, we want to drive three shifts:

• From operator development relying on individual engineering experience → to an operator-driven automated pipeline;

• From domestic chip software ecosystems being unreachably high-bar → to something ordinary engineers can pick up;

• From “having hardware but lacking software” → to tools that help domestic chips truly release their performance and break the bottleneck.

Simply put, KernelCAT is the key booster pushing China’s domestic AI ecosystem from “usable” to “easy to use.”

▐ Making Domestic Chip Migration Less Discouraging, So Developers Focus on Innovation

GBA Voice:​ What scenarios does it cover, and what pain points does it solve?

Luo:​ The most common problem AI developers hit today is: hardware changes, and the model either won’t run or runs poorly. KernelCAT mainly addresses three scenarios:

First, model migration.​ When a model trained on foreign compute platforms like NVIDIA is moved to domestic chips such as Huawei Ascend, you often get complex environment configuration, dependency errors, and unsupported functions in official operator libraries. KernelCAT auto-configures the runtime environment, fills missing dependencies, and even migrates or generates new operators so the model achieves “one-click adaptation.”

Second, performance optimization.​ After migration, models can be slow or power-hungry because certain operators are inefficient. For example, a model that used to do 1,000 inferences per second might drop to 300. KernelADT quickly locates which operator is dragging performance—say, a matrix computation with mis-tuned parameters—and auto-optimizes it, making the model faster and more energy-efficient.

Third, operator development.​ This is the most bottom-level work. Traditionally it requires people who understand algorithms, chip architecture, and compilation optimization all at once—few have that after just a couple years. KernelCAT uses a large model to understand the user’s need. If an engineer says “I need an operator for image feature processing,” it auto-generates code, compressing weeks of development to hours or minutes.

These three scenarios cover the key chain from lab to industrial deployment. KernelCAT’s role is to stop developers from being dragged down by bottom-layer engineering, so they can focus on model innovation itself.

▐ Letting AI Do the Grunt Work: Latency −22%, Throughput +30%

GBA Voice:​ Operator development has long faced the “three highs”—high barrier, high cost, long cycle. How does KernelCAT break this?

Luo:​ “Three highs” shows up like this:

• High barrier, because developers must simultaneously know algorithms, chip architecture, and compilation optimization;

• High cost, because every new chip or new model architecture means re-development;

• Long cycle, because manual tuning takes weeks and can’t keep up with hardware iteration.

KernelCAT’s approach is essentially using AI intelligence to replace human experience, combining the comprehension ability of large models with the mathematical solving power of operations research optimization. The user states the need in natural language—e.g., “I need to optimize the latency of such-and-such operator on Ascend”—and KernelCAT automatically analyzes hardware constraints, decomposes algorithm logic, turns the problem into a mathematical optimization model, then uses algorithms to quickly find the optimal solution. It tests and verifies on the real chip, and outputs ready-to-use operator code.

Typical case:​ On a certain Ascend chip, optimizing one operator traditionally meant tuning a dozen or twenty parameters, each with multiple choices—engineers would try for a long time. KernelCAT found the optimal configuration in just a dozen iterations: latency down 22%, throughput up 30%.

But to fundamentally solve the “three highs,” a tool alone isn’t enough. We need three forces combined:

1. Chip vendors continuously improving hardware performance;

2. The industry co-building open operator libraries and toolchains instead of reinventing wheels;

3. Universities and enterprises cultivating more bottom-layer talent, with policy support in funding and projects so the ecosystem enters a healthy loop.

▐ A Combined Punch: The More People Use It, the Better the Tool Becomes

GBA Voice:​ Despite huge potential, what pressures will KernelCAT face going forward?

Luo:​ To truly compete with NVIDIA’s CUDA ecosystem, tool innovation is the “sharp knife,” but we also need a “combined punch.”

First, hardware must keep up—compute, energy efficiency, stability—especially architectures optimized for large AI models, e.g. better at large-scale data and mixed-precision.

Second, software tooling must be complete. Beyond operator tools, we need training frameworks, inference engines, debuggers—a one-stop environment that feels smooth.

Third, the ecosystem must be active. Enough developers must be willing to use it and contribute code. Policies can encourage students and engineers to join, creating a positive loop: the more people use it, the better the tool gets.

Fourth, policy support. Compute sovereignty is part of national security. The state needs to coordinate at the AI strategy level, embed domestic ecosystem building into science-tech planning, and use funding and tax policy to help the early ecosystem survive “zero to one.”

CUDA took over a decade to build. Developers are used to it. To make them migrate, the key is to let them discover: using domestic chips and tools is fine—performance keeps up, and development is simple. The value of tools like KernelCAT is to slowly build that trust and confidence.

▐ From “Following” to “Leading”: How?

GBA Voice:​ What are the necessary conditions for a mature domestic AI ecosystem that can rival CUDA internationally? Your suggestions?

Luo:​ To break NVIDIA’s first-mover monopoly, several things must happen:

1. Talent.​ Universities and enterprises should cultivate compound talents who understand both chips and algorithms. Let students use domestic AI chips and tools while still in school, step on the potholes early, and mature the ecosystem from the ground up.

2. Tool breakthroughs.​ Tools like KernelCAT must become smarter and more efficient, lowering the usage threshold so users don’t need to learn complex tech to use domestic chips well.

3. Open ecosystem.​ Chip vendors, tool developers, and application companies—autonomous driving, smart manufacturing, etc.—must cooperate and share rather than build silos, forming an open collaborative ecosystem.

4. Policy backing.​ At the critical breakthrough stage, the state should lend a hand—e.g., support early adoption of domestic chips in defense, finance, and other key sectors to help the ecosystem survive its hard early days.

So my suggestion: short term, break through on tools so users want to use it; medium term, refine the ecosystem loop so they use it well; long term, rely on talent cultivation and open collaboration. If upstream and downstream pull together, I believe China’s domestic AI ecosystem can absolutely move from “following” to “leading.”

Source: CMG Greater Bay Area Voice 《科创梦工厂》