中国初创公司Houmo Intelligence在2025世界人工智能大会上发布了专为大型模型设计的全新高效AI芯片Manjie M50，标志着其从智能驾驶向下一代边缘计算的战略转型。该芯片是国内首款将内存与计算集成于单一架构的AI芯片，专为边缘侧的大规模推理优化。M50拥有160 TOPS的INT8算力和100 TFLOPS的bFP16性能，内存高达48GB，带宽153.6 GB/s，能在10瓦功耗下运行15亿至70亿参数的模型，适用于PC、智能音箱、机器人等设备，提供即插即用的AI能力。Houmo CEO吴强表示，M50只是开始，公司愿景是让AI算力像电力一样普及。

💡 **Houmo Manjie M50：国内首款集成内存与计算的边缘AI芯片** Houmo Intelligence在2025 WAIC上推出了Manjie M50，这是中国首款在单一架构中整合内存和计算的AI芯片，专门为边缘设备上的大规模AI模型推理而设计。其强大的性能表现，包括160 TOPS的INT8算力和100 TFLOPS的bFP16性能，以及高达48GB的内存和153.6 GB/s的带宽，使其能够高效运行15亿至70亿参数的模型，同时将功耗控制在10瓦以内，为PC、智能音箱、机器人等终端设备带来强大的本地AI处理能力。

🚀 **战略转型与市场洞察：从智能驾驶到边缘AI** Houmo公司成立于2020年，最初专注于智能驾驶AI芯片。然而，公司CEO吴强在2023年底意识到智能驾驶领域竞争激烈且发展停滞，尤其是对L3级别自动驾驶的信心不足。因此，Houmo果断将重心转向计算-内存集成（CIM）技术，该技术通过将计算单元嵌入内存阵列，有效解决了传统冯·诺依曼架构中的数据移动瓶颈和能耗问题，特别适合处理大型AI模型。这一战略性调整使公司能够抓住边缘AI快速增长的市场机遇。

🔒 **技术迭代与未来展望：DRAM-PIM的潜力** Houmo在不到一年的时间内快速迭代了其产品，推出了M30芯片，并在2024年初成功实现了对600亿参数模型的支持，获得了中国移动的战略投资。公司并未止步于此，正积极研发下一代DRAM-PIM（处理-内存集成）AI芯片，预计2026年面世，目标是实现当前三倍的能效，并支持数十亿参数模型在平板、PC等设备上的广泛本地部署。DRAM-PIM技术有望进一步增强AI芯片的效率和可扩展性，实现真正的实时边缘智能。

📈 **中国边缘AI市场潜力巨大** 随着生成式AI的不断演进，中国正将边缘AI置于国家战略的重点位置。市场数据显示，中国计算-内存集成芯片市场预计到2030年将超过1100亿元人民币。Houmo的成功转型和技术创新，不仅抓住了这一新兴市场机遇，也为中国在边缘AI领域的发展贡献了重要力量，其投资方包括红杉中国、启明创投、源码资本等知名机构，显示了资本市场对公司前景的高度认可。

AsianFin -- China’s Houmo Intelligence unveiled a new high-efficiency AI chip tailored for large models on edge devices at the 2025 World Artificial Intelligence Conference (WAIC), as the startup sharpens its pivot from intelligent driving to next-gen edge computing amid intensifying demand for localized AI inference.

The Houmo Manjie M50, launched in Shanghai during the conference, is China’s first AI chip to integrate memory and computation in a single architecture specifically designed for large-scale inference at the edge. With 160 TOPS of INT8 compute and 100 TFLOPS of bFP16 performance, plus up to 48GB of memory and a bandwidth of 153.6 GB/s, the M50 runs models from 1.5 billion to 70 billion parameters—all within 10 watts of power. The chip is aimed at devices such as PCs, smart speakers, and robots, offering plug-and-play large model capabilities.

“The M50 is just the beginning,” said CEO Wu Qiang at a media briefing. “Our vision is to make AI computing power as accessible as electricity—embedded in every device, across every industry.”

The company also rolled out complementary products including the Liqing-series M.2 cards, Limou-series accelerator cards, and compute boxes, targeting everything from consumer devices to intelligent industrial terminals. It’s the latest sign that China’s edge AI race is heating up as generative models move off the cloud and into everyday products.

Founded in 2020, Houmo initially focused on AI chips for intelligent driving. But by late 2023, Wu concluded the sector was overcrowded and stagnating. “The industry was obsessed with cost competition, and no one believed in L3 autonomy anymore,” Wu said. “We had a chip with strong performance, but the market didn’t want it.”

Instead, the company saw promise in compute-in-memory (CIM)—a chip design that breaks the traditional von Neumann architecture by embedding computation directly into memory arrays. This reduces data movement and energy consumption, addressing bottlenecks in bandwidth and latency—particularly relevant for large AI models.

Houmo re-engineered its first-generation product in under a year, launching the M30 chip for edge large model inference in early 2024. A key vote of confidence came from China Mobile, which used the chip to run a 60-billion-parameter model. In July, the company secured strategic funding from China Mobile’s digital economy funds in Beijing and Shanghai.

“People questioned why I pivoted,” Wu said. “But survival outweighed pride. Autonomous driving was a dead end. Edge AI is a new frontier—and there’s still space to lead.”

Houmo’s new accelerator lineup includes the Limou LM5050 and LM5070 cards, equipped with two and four M50 chips respectively, offering up to 640 TOPS of performance for ultra-large model inference. The company’s transition from SRAM-based CIM to DRAM-based PIM (processing-in-memory) further boosts its hardware’s efficiency and scalability.

Wu says the firm is already developing its next-generation DRAM-PIM AI chip, expected as early as 2026. The chip aims to triple current energy efficiency and support widespread local deployment of models with tens of billions of parameters on devices like tablets and PCs.

“DRAM-PIM is the next step,” said Wu. “It tightens the bond between memory and compute, unlocking real-time intelligence at the edge.”

The WAIC 2025 event, themed “Intelligent Era, Shared Future,” featured more than 1,200 global guests—including 12 Turing and Nobel laureates—and showcased over 3,000 innovations. The exhibition space surpassed 70,000 square meters for the first time, drawing more than 800 companies and revealing more than 100 world or China-first product debuts.

As generative AI evolves, China is increasingly positioning edge AI as a key national focus. Public data suggests China’s computing-in-memory chip market could exceed 110 billion yuan ($15.2 billion) by 2030.

Houmo’s investors include Sequoia Capital China, Qiming Venture Partners, Matrix Partners China, Lenovo Capital, Walden International, and China Mobile.

“Edge intelligence is where the future is headed,” Wu said. “We’re not just building chips—we’re building the infrastructure for the intelligent era.”

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App