Suanmiao 3D TokenPU chip officially enters tape-out phase

🔑 Enhanced Key Takeaways

•Suanmiao Technology has secured nearly RMB 1 billion in Pre-A and Pre-A1 funding rounds, earmarked for the research, development, and mass production of its 3D AI inference chips.
•The company was established in November 2022 by founder Wang Fuquan, a former researcher at the Chinese Academy of Sciences and a key contributor to the Loongson (Godson) CPU project.
•Pre-silicon simulation data for Suanmiao's A4 chip, which utilizes the 3D TokenPU architecture, indicates an inference throughput 1.26x to 2.19x higher than Nvidia H200 on Llama and Mixtral models.
•Suanmiao Technology's core objective is to overcome the 'memory wall' bottleneck in AI computing through a combination of innovative architectural design and the cultivation of a domestic 3D IC supply chain.
•The leadership team includes CTO Liu Ming, who brings over six years of experience in 3D ICs from his time at Loongson, and Chief Scientist Lou Jianguang, a former Microsoft Research principal who collaborated with OpenAI on Excel NLP features and joined in September 2025.

🛠️ Technical Deep Dive

The chip employs a "3D TokenPU architecture" and is referred to as the "A4" chip.
Its design specifically targets large model inference, aiming to mitigate the "memory wall" bottleneck.
The underlying concept of a "Token Processor" involves a hardware-native intelligent accelerator with a minimal software layer, designed for a "Token in → Token out" process with a fixed compute fabric and reconfigurable model parameters/topology.
Technical paths for Token Processors, such as "Token Streaming (Groq route)," suggest a single-core architecture with hundreds of megabytes of on-chip SRAM, allowing model weights to reside directly on the chip to eliminate DRAM bottlenecks.
This approach also involves a compiler that statically orchestrates data paths to achieve deterministic execution, reducing scheduling jitter common in GPUs.
The 3D aspect of the architecture focuses on vertical integration to shorten data travel distances and enhance bandwidth between computing units and memory.
The chip is part of an initiative to produce "100% domestically produced 3D AI inference chips" and leverage a "domestic 3D IC supply chain."

🔮 Future ImplicationsAI analysis grounded in cited sources

China's domestic AI cloud computing hardware market will experience intensified competition and reduced reliance on foreign technology.

Suanmiao's successful tape-out of a high-performance 3D AI chip, backed by substantial domestic funding and a focus on local supply chains, directly supports China's strategic goals for technological self-sufficiency in AI hardware.

Specialized 3D chip architectures will become increasingly crucial for the efficient processing of future large language models and other demanding AI workloads.

The 3D TokenPU architecture is explicitly designed for large model inference and aims to overcome the 'memory wall,' indicating its potential to deliver significant performance improvements for complex AI applications.

⏳ Timeline

2009

Founder Wang Fuquan launched Shengsheng Technology.

2022-11

Suanmiao Technology was founded.

2025-09

Lou Jianguang joined Suanmiao Technology as Chief Scientist.

2026-02

Suanmiao Technology closed Pre-A and Pre-A1 funding rounds, raising nearly RMB 1 billion.

2026-06-17

Suanmiao 3D TokenPU chip officially entered the tape-out phase.

Suanmiao 3D TokenPU chip officially enters tape-out phase

⚡ 30-Second TL;DR

🧠 Deep Insight

🔑 Enhanced Key Takeaways

🛠️ Technical Deep Dive

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline

📎 Sources (6)

👉Related Updates

Xreal Aura XR glasses reservations open at $1,500

Nvidia re-releases RTX 3060 GPU in US market

HyperDimension's Path to Physical AGI

LeDong Robot targets physical AI infrastructure