Orient: Solving the memory bottleneck from the bottom up, gradually advancing the industry.

Orient released a research report, stating that on April 22, 2026, Anker Innovations Technology (300866.SZ) displayed its first neural network integrated AI audio chip with storage and computation capabilities - Thus. Storage-in-computation solves the memory wall bottleneck from the bottom architecture level, optimizing computational power consumption. Currently, there are multiple parallel technical paths for storage-in-computation, moving towards maturity, and domestic players are advancing towards commercialization in AI at the edge and low-power consumption directions. Key points from Orient: Event On April 22, 2026, Anker Innovations Technology showcased its first neural network integrated AI audio chip - Thus. Storage-in-computation addresses the memory wall bottleneck at the basic architecture level, optimizing computational power consumption Some investors may not understand the potential application of storage-in-computation innovation architecture in AI computing. The traditional separated storage and computation architecture proposed by von Neumann in 1945 core feature is the separation of the computation unit and storage unit, leading to frequent data movement between the processor and memory. However, with the rapid development of large-scale model technology, the demand for storage capacity and bandwidth has exponentially increased, significantly amplifying the shortcomings of the separated storage and computation architecture. Based on this, academia and the industry are currently continuously promoting the application of storage and computation integrated architecture. Currently, storage and computation integrated technologies include near-storage computation (integrating computation units with memory through advanced packaging technology), in-storage processing (adding computation functions in the peripheral circuits of memory chips), among others. Among them, in-storage computation is the most integrated solution, directly utilizing the physical characteristics of storage media to perform computational operations inside the storage array, achieving highly parallel and ultra-low power consumption calculations. Compared to traditional solutions, in-storage computation has significant advantages in power consumption, computational efficiency, etc. Under the same process technology, in-storage computation chips are expected to provide higher computing power and lower power consumption per unit area. For example, the WTM2101 chip released by Zhicun Technology focuses on low-power consumption speech interaction scenarios on the edge, with a power consumption of only 5mW, which can increase computing power by 10 to 200 times compared to NPU, DSP, MCU computation platforms at the same power level. Multiple technical paths for storage-in-computation are progressing in parallel, and the technology is moving towards maturity On the storage medium side, the mainstream technical paths for in-storage computation include SRAM, DRAM, Flash, and new resistive random access memory (ReRAM, MRAM, PCM, etc.). Among them, the SRAM storage and computational integrated solution is based on CMOS technology, can use advanced process nodes, and has fast read and write speeds; the DRAM solution has higher storage density than SRAM, suitable for processing large capacity model scenarios; the Flash solution has advantages in non-volatility and low power consumption. Currently, solutions using SRAM, NORFlash, and DRAM as in-storage computation media have had products landing success. New storage media such as RRAM (resistive random access memory), MRAM (magnetic random access memory), PCM (phase-change memory) have good process scalability and ultra-low power consumption characteristics, and are believed to have great application potential. At ISSCC 2026, a joint team from Tsinghua University, Huawei, and ByteDance released a paper on in-storage computation chips, which, for the first time, proposed a hybrid in-storage computation chip based on 28nm process technology. This chip uses RRAM as the storage medium and significantly improves the efficiency and energy efficiency of core operations in recommendation systems through innovative architectural design. The industrialization of in-storage computation is gradually progressing In the direction of AI at the edge and low power consumption, domestic players are advancing towards commercialization. Actions Technology was the first to achieve commercial application of in-storage computation technology in the industry, officially launching AI audio chips for edge scenarios. In 2025, based on self-developed mixed analog SRAM in-storage computation architecture, the shipments of edge AI audio chip products ATS323X and ATS362X from Actions Technology continued to increase. The ATS323X chip has quickly landed in flagship wireless microphones of brand customers and been successfully released on the market, penetrating multiple professional audio leading brand supply chains. Zhicun Technology's WTM2101 chip, the world's first in-storage computation integrated speech chip based on NORFlash, has achieved shipments of over 10 million, applied in smart wearable devices of brands such as Huawei and Xiaomi. Anker Innovations Technology's Thus chip, built on NORFlash technology, natively supports a 4 trillion parameter model and will be deployed in the flagship headphones Anker is about to release. With 3D-CIM technology architecture, WinnowCore has deep cooperation with domestic leading memory manufacturers and multiple terminal industry leaders and has been designated as the leading unit for the RISC-V in-storage computation application group, launching the world's first RISC-V in-storage computation standard development work in Xiaoshan, Hangzhou; the company received an investment from GigaDevice Semiconductor Inc. in March 2026. Investment recommendations and investment targets Storage-in-computation addresses the memory wall bottleneck from the bottom layer, and the industry is gradually advancing. Related investment targets: GigaDevice Semiconductor Inc., Actions Technology, Zbit Semiconductor, Inc., Anker Innovations Technology, etc. Risk warning AI landing falls short of expectations, technology iteration speed falls short of expectations, domestication progress falls short of expectations.