Can Reconfigurable AI Chips Break the Compute Wall? Inside China's RPU Revolution

On September 4, the A‑share market witnessed a dramatic drop as Cambricon closed at 1,202 CNY, a more than 20% pull‑back from its all‑time high of 1,595.88 CNY set on August 28, erasing over 700 billion CNY in market value in a single day. Around the same time, Nvidia announced on September 3 that its H100/H200 chips have ample inventory, but the China‑specific H20 chips faced "security controversies" and missed first‑half sales targets, highlighting the complex AI compute competition and the three technical challenges of efficiency, interconnect, and storage walls.

A Chinese chip startup, Qingwei Smart, recently praised by the People's Daily, is developing an original reconfigurable AI chip (RPU) to create a differentiated path in the domestic chip market.

Architecture Innovation: How Reconfigurable Computing Solves Compute Bottlenecks

Reconfigurable AI chips (RPU) represent a distinct AI chip paradigm, often described as a "data‑flow architecture". Their principle resembles a railway switch: fixed tracks of traditional chips are replaced by dynamically configurable compute units, allowing the chip to adapt instantly to tasks such as speech recognition, image analysis, and inference.

This dynamic reconfiguration gives RPUs superior compute efficiency, scalability, concurrency, and cost‑performance compared to GPUs, earning them the nickname "Transformers of the chip world". The International Technology Roadmap for Semiconductors (ITRS) lists reconfigurable chips as the most promising future architecture, positioned after CPUs, FPGAs, and GPUs as the fourth class of general‑purpose compute chips.

Huatai Securities' overseas technology chief, He Pianpian, notes that the long‑term value of reconfigurable chips lies in delivering exceptional performance and efficiency for specific AI workloads. Their success will depend on building a robust software ecosystem to attract developers away from existing GPU ecosystems.

Commercial Breakthrough: Accelerating AI Applications

Last year, Qingwei mass‑produced the cloud‑centric TX81 chip. At a comparable 1,000 P compute scale, TX81 demonstrated stronger interconnectivity and energy efficiency than GPU clusters. Since its launch, the product has been deployed in multiple domestic kilowatt‑scale AI compute centers, accumulating orders for nearly 20,000 units.

Global Landscape of Data‑Flow Architectures

Worldwide, data‑flow architectures are gaining momentum. In July, OpenAI began renting Google TPUs for ChatGPT inference, signaling diversification of AI chip architectures. Stanford‑spun SambaNova has become an AI‑chip unicorn with its own reconfigurable chip, while U.S. startup Groq’s tensor‑flow processor (LPU) offers ten‑fold faster inference at one‑tenth the cost of Nvidia GPUs.

Frontier Chip Technologies and Wafer‑Level Integration

The AI compute industry faces exponential growth in model parameters and the physical limits of Moore's Law. Wafer‑level chip technologies are seen as a key path to break spatial constraints, dubbed the "star of tomorrow". Huatai’s analysis points out that as AI chip counts rise, inter‑chip bandwidth becomes the primary bottleneck, which wafer‑level integration can directly address.

Qingwei’s C2C compute‑grid technology, based on reconfigurable data‑flow, enables point‑to‑point direct connections between multiple chips, avoiding the bandwidth and latency limits of traditional switch‑based GPU interconnects.

Industry Perspectives

Qingwei CEO Wang Bo told the People’s Daily that global AI compute demand is surging and reliance on a single vendor or technology path limits competitive evolution. China must simultaneously develop both major AI chip streams to achieve true autonomy.

Overall, the AI compute arena resembles a martial‑arts tournament where GPU and data‑flow schools vie for supremacy; emerging players like Google’s TPU and Qingwei’s RPU are climbing to the "summit" and could drive a healthier, more sustainable AI industry worldwide.

Disclaimer: This article aggregates information from official media and online sources. It does not constitute investment advice; readers should verify details with the latest data.