The pot of tea in Zeng Yi’s office was refilled several times, and the conversation unwound at that pace, in the manner typical of Guangdong hospitality. Nothing formal, nothing vague either. He speaks with the residue of a Chaoshan accent, unhurried, finishing his sentences with a smile that narrows his eyes and tips him slightly forward.
Talking about AI and RISC-V, and where the two meet in the industry, he cannot quite contain his enthusiasm.
Over the past two or three years RISC-V has moved faster in China than he expected. Edge and internet of things applications have rolled out comprehensively, and in Shenzhen and the Greater Bay Area in particular a large number of companies are switching from ARM. What excites him more is the variable AI has introduced, because every major technology group is now using RISC-V to build customised AI processors, Qualcomm, Google and Meta included.
In his reading, this is the first time RISC-V has touched a genuinely large opportunity.
Sunzhan sits on that line. Founded in 2022 with around twenty people, it bets on domain-specific architecture, supplying customers with customised CPU, DSP and NPU cores. Zeng spent twenty years in the industry and knows exactly how poorly general-purpose processors serve the demands of chips and algorithms in the post-Moore era.
Three years on, the business is genuinely running. As customers have grown closer to the company, the obvious awkward question is whether they worry about being tied to the Sunzhan platform. He did not dodge it. The whole point, he said with a laugh, is to get customers aboard our pirate ship. Then he added that what actually keeps them is not lock-in but the value customisation delivers.
That is the most distinctive thing about him. He does not avoid the commercial reality of the business, and he does not disguise what he wants. He has little faith in trends conjured out of nothing, and much more in paths that demand forces into existence. ARM, as he repeats several times, was forced out by the low-power requirements of mobile phones. This time his bet is that AI will force RISC-V into the foreground.
Who decides the architecture
Asked what AI has changed at the foundational level, given how stable processor architecture competition was for years, Zeng does not hedge. AI has changed the landscape enormously, he says, and arguably shattered it. The design paradigm of the whole chip industry is shifting, or more precisely the power to decide chip architecture is moving. It used to sit with chip companies, and everyone else wrote programmes to suit them. Now it sits with AI companies, because the model decides the architecture.
Customers used to run all kinds of algorithms and applications on general-purpose CPUs from Intel, which made the CPU’s position very secure. Different AI models now have completely different compute requirements, and a single general-purpose CPU can no longer support them efficiently. Nvidia GPUs covered the early phase, but as AI moves to the edge and lands inside all manner of intelligent hardware, a device usually needs a dedicated chip to run a particular model.
At the microarchitecture level the role of the CPU is weakening. Faced with AI models, the vector processing unit and the tensor or neural processing unit matter more. Coordination between them, and architecture and instruction set optimisation aimed at specific models and algorithms, has become the core requirement of the AI era. That is why Intel has been hit noticeably and why ARM’s importance has relatively declined.
RISC-V stands out at this moment precisely because all of those processing units need to be customised heavily around models. Different models need different custom instructions, and those instructions either do not exist on closed architectures such as ARM and Intel or run at very low efficiency. So companies have started developing on RISC-V themselves.
The trend is now industry consensus. Qualcomm and Tesla have each acquired RISC-V companies. Google, Meta and Intel itself have all established deep partnerships with RISC-V firms. Without exception, these arrangements use the open, flexible and freely modifiable RISC-V architecture as the base for next-generation AI processors.
The key advantage over x86 and ARM, he confirms, is customisability and open source. That openness lets each company modify and optimise the architecture deeply for its own needs, and the freedom to rework it at that level is exactly where RISC-V’s superiority shows. Closed architectures where the vendor controls extension rights are, in his words, clearly no longer enough. The industry has settled on the view that simply buying an ARM CPU to run an AI model belongs to the past, replaced by dedicated designs integrating several kinds of processing unit.
Design as a platform
On where AI catalyses RISC-V most, his answer is not what one might expect. The biggest catalytic effect, he says, is that AI has turned chip design technology into a platform.
Five to ten years ago a small company getting into CPU design was almost unimaginable. Today CPU design is everywhere. Sunzhan and a partner brought out a new CPU in two months, which was only possible because of that shift. AI is used heavily inside the company’s own engine, generating design proposals quickly, and it has recently been pushed further into the code generation engine. He tells colleagues internally that Sunzhan should become the coding assistant of the processor industry, using AI to help users generate domain-specific architecture processors automatically.
AI has knocked down what used to be a forbiddingly high technical barrier in chip design, he says, and in doing so created a large number of market opportunities.
Markets decide architectures
Asked what really determines whether an architecture matures and gets adopted widely, Zeng calls the question a big one and answers with market demand.
We treat the von Neumann architecture as canonical, he points out, but academically it is not the most efficient option. It won largely thanks to military demand after the Second World War and widespread adoption in the United States, rather than pure technical superiority. Equally, ARM was not inherently the best choice for phones. It is a RISC architecture like any other, and what mattered was that phones are extremely sensitive to power consumption and the market urgently needed something other than Intel’s approach. That intense demand for low power is what made ARM.
RISC-V follows the same rule. It has been pushed forward by the AI wave, by demand for customisation and by the trend towards processor specialisation. Discussing architecture in isolation from the market means little. Academia has accumulated a large stock of processor architectures, but they only get reactivated and refined when the market produces a clear and forceful demand.
Domestically, he sees two paths. One targets servers, loud but slow to commercialise. The other targets domain-specific architecture and reaches commercial deployment faster. More and more companies are moving into the domain-specific direction, which tells you the market is making its choice. Ecosystem building remains indispensable, and China is working through it steadily. It takes time, but the direction is settled.
The breakout point
On how long a stable three-way split with ARM and x86 might take, Zeng starts from the current map. Intel dominates servers and personal computers, ARM dominates phones, and ARM has spent years trying to move up into personal computers and servers without an easy time of it. By comparison RISC-V has moved quickly, only really spreading after the pandemic in 2021 and 2022, and it has since caught fire in edge and internet of things applications, especially in Shenzhen and the Greater Bay Area, where many companies have adopted it or are considering it.
A genuine three-way split needs a breakout point, and he doubts it will come from a frontal assault on established markets. ARM did not rise by swallowing Intel’s territory. It seized a completely new opportunity in mobile phones. RISC-V’s opportunity lies in catching the AI wave, and the likely path is that companies use it to build AI processors of every shape, then work back into servers, personal computers and phones through AI applications.
On whether RISC-V will become a major server architecture, he says he lacks the information to judge, and that the question carries an assumption he wants to challenge. People asking it have already decided servers are a good market, so they want to know when Intel gets overtaken. His own doubt is whether that market is as attractive as everyone thinks.
Sunzhan considered entering multicore and server CPUs early on and built technical reserves for it. Since last year, the arrival of the data processing unit has convinced him the market is changing and may become considerably less appealing. Putting resources into NPUs appears to pay far better. An NPU is a classic domain-specific architecture, sacrificing generality for extreme efficiency on specific tasks, so at the same power and area its inference throughput is far above a general-purpose RISC-V CPU.
Perhaps in future the CPU really will matter less, he adds, and computing devices may no longer depend on traditional operating systems such as Windows or Android. The question genuinely worth thinking about is what a processor should be doing when the AI era arrives.
The right ratio
Put to him that the token economy has brought CPUs back into focus, with some in the industry suggesting a future CPU, GPU and NPU ratio of one to one to one, Zeng accepts the view exists and says he does not believe CPUs lose their place. His position is not to fix on one slice of the CPU market but to return to what the market needs.
What it needs is a complete AI solution, and his judgement on the composition is clear. It is a combination of CPU, VPU and NPU. Sunzhan has products across all three and offers customisation, with the ratio adjusted flexibly to the actual application.
The company is therefore doing two things: building a rich enough shelf of products, and integrating them into processor solutions fitted closely to a customer’s model and vertical domain. Such a solution typically contains several CPUs and VPUs and a larger number of NPUs. Within that system, he stresses, the CPU remains vital.
On high-performance general computing, which RISC-V cannot avoid if it is to become a genuine third architecture, he answers in layers. At the architecture and standards level, RISC-V already has the technical foundation for high-performance computing. Whether x86, ARM or RISC-V, any of them can produce top-standard processor cores and architectures given enough development. At the implementation level, driven in particular by policy on domestic capability, sustained investment from institutions including the Institute of Computing Technology at the Chinese Academy of Sciences and Alibaba, and the progress of projects such as XiangShan and Kunminghu, technical feasibility is established and the gap is narrowing.
Commercially it is more complicated. The difficulty lies less in the technology than in the ecosystem and market acceptance, because the server field carries an enormous accumulated software ecosystem, and hitting the performance target is only the beginning.
Editor’s note: this English report is an adapted translation of a Chinese-language original published by LeiPhone (leiphone.com). Figures, dates and direct quotations follow the source.