Author: Su Ziyun | Publication Date: April 16, 2026, 12:13 PM
According to Wccftech, Qualcomm is collaborating with Chinese memory manufacturers Changxin Memory and GigaDevice to develop a standalone neural processing unit (NPU) that incorporates 3D DRAM, attempting to overcome the current bottleneck of mobile AI computing power being limited by the SoC architecture.
This solution aims to enhance long-term computing capabilities such as real-time translation and image generation by separating AI computing from the main chip and importing it into high-bandwidth memory. It is expected to be introduced into the mid-to-high-end mobile phone market in China as early as the end of 2026 to the beginning of 2027.
Currently, most smartphones integrate their NPUs into the SoC, which requires sharing power consumption and memory resources with the CPU and GPU, making it difficult for AI computing power to be output stably for extended periods.
The new solution adopts an independent architecture. According to Ming-Chi Kuo, an analyst at TF International Securities, it may be equipped with 4GB of customized 3D DRAM and significantly improve memory bandwidth through advanced packaging technologies such as TSV (Through Silicon Via) and Hybrid Bonding, surpassing the current LPDDR5X standard, thereby supporting more intensive AI tasks.
In terms of performance, this AI accelerator can provide a stable computing power of approximately 40 TOPS. In comparison, while current flagship phones boast a maximum of 100 TOPS, most can only achieve this briefly under specific conditions. If stable output can be maintained for an extended period, it has the potential to substantially improve the AI experience on mobile devices and even double the overall computing power.
However, cost pressures could be a key variable in its adoption. With rising memory prices, Chinese brands such as Xiaomi, OPPO, vivo, and Huawei are actively seeking to reduce DRAM and storage costs to maintain product profit margins. In this context, adding an additional dedicated NPU with 4GB of 3D DRAM will inevitably increase the overall BOM (Bill of Materials), potentially impacting manufacturers' willingness to adopt it.
Furthermore, the market demand for mobile AI remains unclear. TF International Securities analyst Ming-Chi Kuo points out that even with gradually maturing hardware, there is still a lack of application scenarios that can fully leverage the AI capabilities of devices. Consumers remain hesitant about paying a premium for this. While the combination of a dedicated NPU and 3D DRAM has the potential to reshape mobile AI architecture, its ultimate mainstream adoption depends on cost control and the maturity of the application ecosystem.
Furthermore, the market demand for mobile AI remains unclear. Ming-Chi Kuo pointed out that even as hardware conditions gradually mature, there is still a lack of application scenarios that can fully leverage the AI capabilities of devices, and consumers are still taking a wait-and-see approach regarding whether they are willing to pay a premium for it.
technews.tw
According to Wccftech, Qualcomm is collaborating with Chinese memory manufacturers Changxin Memory and GigaDevice to develop a standalone neural processing unit (NPU) that incorporates 3D DRAM, attempting to overcome the current bottleneck of mobile AI computing power being limited by the SoC architecture.
This solution aims to enhance long-term computing capabilities such as real-time translation and image generation by separating AI computing from the main chip and importing it into high-bandwidth memory. It is expected to be introduced into the mid-to-high-end mobile phone market in China as early as the end of 2026 to the beginning of 2027.
Currently, most smartphones integrate their NPUs into the SoC, which requires sharing power consumption and memory resources with the CPU and GPU, making it difficult for AI computing power to be output stably for extended periods.
The new solution adopts an independent architecture. According to Ming-Chi Kuo, an analyst at TF International Securities, it may be equipped with 4GB of customized 3D DRAM and significantly improve memory bandwidth through advanced packaging technologies such as TSV (Through Silicon Via) and Hybrid Bonding, surpassing the current LPDDR5X standard, thereby supporting more intensive AI tasks.
In terms of performance, this AI accelerator can provide a stable computing power of approximately 40 TOPS. In comparison, while current flagship phones boast a maximum of 100 TOPS, most can only achieve this briefly under specific conditions. If stable output can be maintained for an extended period, it has the potential to substantially improve the AI experience on mobile devices and even double the overall computing power.
However, cost pressures could be a key variable in its adoption. With rising memory prices, Chinese brands such as Xiaomi, OPPO, vivo, and Huawei are actively seeking to reduce DRAM and storage costs to maintain product profit margins. In this context, adding an additional dedicated NPU with 4GB of 3D DRAM will inevitably increase the overall BOM (Bill of Materials), potentially impacting manufacturers' willingness to adopt it.
Furthermore, the market demand for mobile AI remains unclear. TF International Securities analyst Ming-Chi Kuo points out that even with gradually maturing hardware, there is still a lack of application scenarios that can fully leverage the AI capabilities of devices. Consumers remain hesitant about paying a premium for this. While the combination of a dedicated NPU and 3D DRAM has the potential to reshape mobile AI architecture, its ultimate mainstream adoption depends on cost control and the maturity of the application ecosystem.
Furthermore, the market demand for mobile AI remains unclear. Ming-Chi Kuo pointed out that even as hardware conditions gradually mature, there is still a lack of application scenarios that can fully leverage the AI capabilities of devices, and consumers are still taking a wait-and-see approach regarding whether they are willing to pay a premium for it.
高通攜長鑫推獨立 NPU＋3D DRAM，突破手機 AI 性能瓶頸
根據 Wccftech 報導，高通正與中國記憶體廠長鑫存儲及兆易創新合作，開發一款結合 3D DRAM 的獨立神經處理單元（NPU），試圖突破現行手機 AI 算力受限於 SoC 架構的瓶頸。 該方案透過將 AI 運算從主晶片中分離，並導入高頻寬記憶體，目標強化即時翻譯、影像生成等長時間運算能力，最快...
technews.tw
Last edited by a moderator: