Author | Bao Yonggang, Editor | Liu Wei
Leiphone.com has exclusively learned that MemSilicon has completed its Series A funding round of over CNY 400 million. The round was led by Boyuan Capital, with follow-on investments from Bosch Ventures, Chuangling Capital, miHoYo, Zhongguancun Science City Company, GigaDevice Semiconductor, and existing shareholders Furong Investment, CDB Venture Capital, Hengyu Capital, Jinqiu Fund, Shunxi Fund, Xiaomi Strategic Investment, and CAS Star continuing to increase their stakes.
Lu Tao, former President and General Manager of Greater China at Graphcore, has also joined MemSilicon as CEO and Co-founder.
With over 20 years of deep experience in the semiconductor industry, Lu Tao has previously served as General Manager of Cavium China and Vice President of Lightelligence. His expertise spans Arm server CPUs, AI processor IPUs with large SRAM architectures, and silicon photonics.
With Lu Tao's joining, MemSilicon will continue to advance the industrialization and commercialization of Resistive Random Access Memory (RRAM), while launching the R&D of its first-generation LPU, extending its business from memory to computing.
Jensen Huang previously mentioned, "Speed is intelligence." Lu Tao understands this in two dimensions: good memory and a fast brain. In the context of AI, the former refers to context capability, and the latter to inference latency and generation speed.
This resonates with the two strategic paths of memory and computing that MemSilicon is betting on. The company researches not only how to store data but also how to compute and retrieve it faster, ultimately achieving synergy at the system level.
Lu Tao's goal is very clear: to build an LPU that offers "full-parameter models at high-speed rail prices."
The intersection between Lu Tao and MemSilicon began in 2024 when the company was founded. At that time, MemSilicon invited him to serve as CEO, but he did not join. Two years later, what changed his decision?
01. Amid the LPU Boom, Knowing Both How to Do It and Why to Do It
The GTC in March this year became a trigger for Lu Tao's decision. Watching Jensen Huang launch products, he strongly missed the days of competing with chip giants: challenging NVIDIA at Graphcore, competing with Intel on server CPUs at Cavium, and competing with Broadcom on switching chips.
Starting a business from scratch was an option. But he calculated the time required: building a team, finalizing the architecture, and closing the funding round would take six to nine months, even in a hot capital market. He was not sure how long the hype would last.
"Speed is extremely, extremely important." Joining MemSilicon would allow him to get a faster start.
His decision not to join MemSilicon in 2024 was also related to the stage both parties were at. At that time, MemSilicon prioritized moving memory technology from the lab to products, and Lu Tao felt the value he could contribute was limited. Before this current joining, however, MemSilicon had already decided to enter the computing field, where he could independently advance the products he wanted to build.
More importantly, the MemSilicon team originates from Academician Liu Ming's team at the Chinese Academy of Sciences, possessing over 20 years of accumulated expertise in memory technology. In addition to RRAM, the team also has a research foundation in SRAM, MRAM, DRAM, and other fields, which can be combined with his experience in computing architecture and industrialization.
After joining, Lu Tao quickly built a core LPU team of over a dozen people, covering core areas such as hardware, compilers, and system software. He particularly values the practical experience gained from working with top-tier customers at Graphcore in the past—how to implement new architectures, and the reasons behind both successes and failures.
"It is very difficult to build an AI chip, and it is even more challenging to sell it," Lu Tao deeply realized.
He summarizes the team's capabilities with know-how (how to do it) and know-why (why to do it): knowing not only how to build the chip but also why to build it, and what ultimate value to provide to users. In his view, in today's landscape with numerous GPU and NPU companies, new teams need to determine their architecture starting from unsolved problems.
02. LPU Makes High-Speed Inference as Mainstream as High-Speed Rail
The product goal of MemSilicon's first LPU is to target full-parameter models with trillions of parameters, without relying on quantization, speculative decoding, and other techniques, pursuing a speed of generating 1,000 tokens per second per user.
Beyond speed, there is also price.
Lu Tao uses high-speed rail as an analogy: users are willing to pay for better speed and experience, but if a regular train ticket costs CNY 200 while a high-speed rail ticket costs CNY 2,000, it will be hard to popularize. The same applies to high-speed inference; it can be more expensive than regular services, but the premium must be limited.
"It will not be a niche in the future, but a mainstream one." Lu Tao hopes to enable MemSilicon's LPU to run full-parameter models, achieving high-speed tokens and leapfrog cost-effectiveness.
To achieve this goal, the key lies in the joint innovation of memory, computing, and interconnects.
The Graphcore IPU, whose commercialization Lu Tao once participated in, adopts a large SRAM architecture. In his view, SRAM can provide high bandwidth and low latency, but when facing models with trillions of parameters, capacity becomes a constraint. If more chips are added to accommodate the model, cross-chip interconnects will bring additional costs, power consumption, and time overhead.
Therefore, MemSilicon's LPU aims to retain the speed advantage of SRAM while introducing the capacity advantage of other media, using fewer chips to host large models and controlling the interconnect overhead.
"Beyond SRAM, we also have a new memory media layer." Lu Tao introduced the design concept of the first-generation LPU. This design aims to balance speed, capacity, and cost at the system level, while also taking system throughput into account.
Lu Tao clarified that the first-generation LPU will not use RRAM, but will adopt more mature memory media. RRAM still needs to overcome capacity challenges, and efforts will be made to integrate it with computing products in the future.
Currently, MemSilicon is in discussions with top-tier cloud service providers and large model companies. For the next two to three years, Lu Tao hopes to deliver computing products into the hands of customers while simultaneously advancing the R&D of high-capacity RRAM.
MemSilicon's differentiation lies in the combination of memory device and process accumulation with computing architecture, software, and industrialization experience within the same team. Lu Tao's joining gives the company the opportunity to extend its technological accumulation from memory to complete computing systems, also providing room for continuous evolution towards "full-parameter models at high-speed rail prices."
For more information on domestic chip companies, feel free to add the author's WeChat BENSONEIT to exchange insights.