On October 19, Beijing Mingxin Qirui Technology Co., Ltd. (hereinafter referred to as Mingxin Qirui) announced the completion of over RMB 400 million in Series A funding, led by Boyuan Capital, with participation from Bosch Ventures, miHoYo, GigaDevice Semiconductor Inc. (ASX: 603986), and continued follow-on investment from existing shareholder Xiaomi Strategic Investment. The company stated that this round of funds will be used for RRAM (Resistive Random Access Memory) industrialization, commercialization, and the development of the new LPU (Language Processing Unit) product. Relevant reports did not disclose the valuation or equity dilution ratio for this round.
Two additional developments were disclosed on the same day. Former Graphcore President and Greater China General Manager Lu Tao joined the company as CEO and co-founder. Lu Tao previously served as General Manager of Cavium China and Vice President of Lightelligence, with experience in server CPUs, AI processors, and silicon photonics. The company also simultaneously launched the development of its first-generation LPU, targeting large model inference scenarios, with a focus on optimizing Token generation speed, inference latency, and service cost.
This change had prior technological groundwork. On October 14, Mingxin Qirui released tape-out information for its first-generation 1Gb RRAM chip, Taishan No. 1. The company stated that Taishan No. 1 is based on the 1TnR technology route (one selector transistor corresponding to multiple resistive memory cells), specifically a 1T4R structure, meaning one selector transistor corresponds to four resistive memory cells, paired with 28nm CMOS process. The company disclosed a storage density of 30Mb/mm², access latency of less than 20 nanoseconds, and integrated on-chip hardware error correction circuitry. Currently, Taishan No. 1 has completed tape-out and is undergoing verification, with no mass production progress disclosed yet.
Mingxin Qirui was established in May 2024, with core team members from the Chinese Academy of Sciences and Fudan University. The team initially focused on RRAM R&D and industrialization and is now expanding into large model inference chips. In terms of product layout, Mingxin Qirui stated that the company currently has three main product lines. The first is embedded RRAM IP, targeting display drivers, microcontrollers, SSD controllers, and security chips. The company claims it has entered customer scenarios and plans to scale up, without disclosing customer names, order amounts, or shipment volumes. The second is high-capacity RRAM, namely Taishan No. 1, which is still under verification. The third is hybrid compute-in-memory combining RRAM with DRAM, targeting edge inference scenarios such as embodied intelligence and AI PCs, exploring hybrid 3D integration, without disclosing specific chip models, tape-out progress, or commercialization milestones.
Lu Tao publicly stated that the first-generation LPU will not use RRAM but will use more mature storage media, with the product goal of supporting complete models with trillions of parameters and achieving 1,000 Tokens per second per user. This metric is a research and development target and has no measured performance support yet. Mingxin Qirui confirmed to All-Weather Technology that the LPU is currently in the R&D stage, without further disclosing tape-out, verification, or delivery timelines.
From the disclosed information, Mingxin Qirui's technology layout presents three characteristics, with performance and cost advantages still awaiting verification. First, it extends from memory devices to inference systems. While advancing RRAM IP and standalone memory, the company has launched LPU R&D, attempting to simultaneously master storage media, computing architecture, and system collaboration capabilities. This direction is not unique to Mingxin Qirui; in the industry, companies such as Xinyi Technology currently cover RRAM storage and AI inference chips. Second, 1TnR is a verifiable technical design. The company claims that circuit solutions such as negative voltage design, multi-voltage domain wordline drivers, and on-chip hardware error correction can improve leakage, miswriting, and read reliability. However, relevant advantages still lack third-party test data and same-caliber comparisons with competitors. Third, the LPU route has similar goals to Groq. Currently, Mingxin Qirui has not publicly disclosed its complete architecture, so a strict chip architecture comparison cannot be made.
From the perspective of industrialization progress, companies in the industry have already disclosed mass production or commercialization progress. Xinyi Technology's official website states that its standalone memory achieved mass production in 2021, with cumulative shipments exceeding 30 million units. Yizhu Technology claims to have mass-produced all-digital compute-in-memory chips based on new memory. Houmo AI has launched compute-in-memory AI chip products such as M30 and M50. International manufacturer Groq already provides commercial inference services based on LPU through GroqCloud. In comparison, Mingxin Qirui's Taishan No. 1 is still under verification, and the first-generation LPU is in the early R&D stage, with product delivery and commercialization scale still awaiting observation.
Mingxin Qirui is expanding from RRAM storage to AI inference systems, and subsequent attention is needed on whether technical capabilities can be converted into mass production orders and customer revenue. Material consistency, durability, yield, process compatibility, and software ecosystem are issues that typically need to be addressed during the industrialization of new memory technologies. Worth further observation are the progress of Taishan No. 1 verification completion and customer sampling, whether embedded RRAM IP can disclose specific customer names and order scales, and the first tape-out timeline for the first-generation LPU.
Comments