HiSilicon
HiSilicon Technologies Co., Ltd. is a fabless semiconductor design company wholly owned by Huawei Technologies Co., Ltd., focused on developing application-specific integrated circuits (ASICs), system-on-chips (SoCs), and other semiconductors for telecommunications equipment, mobile devices, and data processing applications.[1][2] Established in 2004 as Shenzhen HiSilicon Semiconductor Co., Ltd. from Huawei's earlier ASIC design center originating in 1991, the company is headquartered in Shenzhen, China, and employs thousands of engineers dedicated to chip architecture innovation.[3][4] HiSilicon's flagship products include the Kirin series of mobile SoCs, which power Huawei smartphones with integrated CPU, GPU, and modem capabilities, and the Ascend series of AI accelerators designed for machine learning workloads in servers and edge computing.[2][5] Notable achievements encompass the Kirin 9000S, a 7-nanometer process chip produced domestically by SMIC for the Huawei Mate 60 Pro in 2023, marking a breakthrough in China's advanced semiconductor fabrication amid international export restrictions.[5] These restrictions, imposed by the United States in 2019 citing national security risks tied to potential military end-use and intellectual property concerns, have compelled HiSilicon to pivot toward indigenous supply chains, fostering advancements in areas like high-bandwidth memory integration for next-generation Ascend processors despite acknowledged performance gaps relative to leading Western designs.[6][7]Corporate Background
Founding and Ownership
HiSilicon Technologies Co., Ltd. (Chinese: 海思半导体有限公司) originated as Huawei Technologies Co., Ltd.'s ASIC Design Center, established in 1991 to develop application-specific integrated circuits for telecommunications equipment.[8][9][3] In October 2004, the unit was formally incorporated as Shenzhen HiSilicon Semiconductor Co., Ltd., transitioning into an independent entity focused on semiconductor design while retaining close integration with Huawei's broader operations.[8][10] This restructuring followed Huawei's early experiences in chip development, including a 2003 intellectual property dispute with Cisco Systems that underscored the need for in-house capabilities.[10] HiSilicon operates as a wholly owned subsidiary of Huawei Technologies Co., Ltd., which was founded in 1987 by Ren Zhengfei as a reseller of telecommunications gear before expanding into manufacturing and R&D.[9][11] Huawei's ownership structure is employee-based through a trade union committee holding shares on behalf of approximately 100,000 staff, enabling long-term strategic investments without external shareholder pressures, though HiSilicon's control remains centralized under Huawei's corporate governance. The subsidiary is headquartered in Shenzhen, China, with its legal name reflecting its focus on HiSilicon-branded technologies derived from "high silicon."[1][11] This ownership model has supported HiSilicon's growth into a fabless semiconductor designer, reliant on foundries like TSMC for manufacturing prior to U.S. export restrictions.[2]Leadership and Organizational Structure
HiSilicon operates as a wholly owned subsidiary of Huawei Technologies Co., Ltd., with its leadership integrated into Huawei's broader rotating chairmanship system, where senior executives from the parent company periodically oversee strategic direction. The subsidiary maintains a dedicated executive team focused on semiconductor design and operations, reporting ultimately to Huawei's board of directors.[12] As of September 2025, Jeffery Gao serves as chairman and legal representative of HiSilicon Semiconductor Co., Ltd., having succeeded Eric Xu Zhijun in the role following Xu's resignation amid Huawei's internal reorganization to bolster chip development. Gao, previously head of HiSilicon's Shanghai division and reportedly the CEO, brings extensive experience in Huawei's router business and regional operations. He Tingbo holds the position of president, a role she assumed after Xu Wenwei's retirement, overseeing R&D and strategic initiatives; she joined Huawei in 1996 as a chip engineer and advanced through positions including chief ASIC engineer, R&D director, and president of the 2012 Laboratories. In July 2025, He was additionally appointed head of Huawei's Senior Talent Compensation Division to support semiconductor talent retention.[13][14][15][16][17] HiSilicon's organizational structure emphasizes functional divisions centered on integrated circuit design, with primary headquarters in Shenzhen and specialized R&D centers in Beijing, Shanghai, Silicon Valley (United States), and Sweden to leverage global talent in processor architectures, modems, and AI technologies. The company employs a hierarchical model typical of fabless semiconductor firms, prioritizing engineering and innovation teams under executive oversight, though detailed internal hierarchies remain proprietary and aligned with Huawei's employee-shareholding system for incentive alignment. Key directors include figures such as Ji Wang, Danny So, and Jian Lin, supporting operational and technical leadership.[18][19]Historical Development
Early Expansion (2004-2018)
HiSilicon Technologies Co., Ltd. was formally established on September 28, 2004, as Shenzhen HiSilicon Semiconductor Co., Ltd., evolving from Huawei's internal ASIC design center founded in 1991, with a focus on fabricating custom semiconductors for telecommunications infrastructure.[20] Initially, the company prioritized ASICs for Huawei's core networking gear, including router processors, baseband modems, and optical network chips, enabling cost efficiencies and customization in Huawei's carrier equipment deployments.[2] [21] By the mid-2000s, HiSilicon broadened its scope beyond telecom ASICs into multimedia applications, developing video codec processors and chips for set-top boxes, digital televisions, and surveillance systems, which supported Huawei's entry into consumer electronics markets.[22] This diversification aligned with Huawei's vertical integration strategy, allowing in-house silicon to optimize performance in emerging products like IPTV solutions and early multimedia devices. The Balong modem series emerged during this period, with initial 3G models paving the way for multi-mode baseband solutions; by 2014, the Balong 711 became one of the earliest 4G modems, certified by over 100 operators and integrated into Huawei handsets for LTE-FDD/TDD, WCDMA, and GSM support.[23] A pivotal shift occurred in 2011 when Huawei, under its Consumer Business Group, committed to proprietary mobile SoCs to lessen reliance on external vendors like Qualcomm, culminating in the 2012 release of the Kirin K3 series—a tri-core ARM-based processor used in devices such as the Huawei Ascend D2.[2] The Kirin lineup formalized in 2014 with the Kirin 910, the world's first commercial quad-core 64-bit mobile SoC on a 28 nm process, powering mid-range phones like the Ascend P7 and tablets. Successive advancements followed: the 2015 Kirin 950 introduced octa-core big.LITTLE architecture on 16 nm; the 2016 Kirin 960 added Vulkan API support and enhanced Mali GPU integration on 16 nm; and the 2017 Kirin 970 debuted a dedicated neural processing unit (NPU) for on-device AI acceleration, fabricated on TSMC's 10 nm process.[2] These SoCs fueled Huawei's smartphone shipment growth, with HiSilicon capturing increasing adoption in premium Android devices. In parallel, HiSilicon advanced modem technology, unveiling the Balong 5G01 in 2018 as the first commercially available 3GPP-compliant 5G chipset, supporting sub-6 GHz and mmWave bands for early 5G trials.[24] By late 2018, Huawei planned expanded use of HiSilicon SoCs across more models, reflecting the subsidiary's maturation into a key enabler of Huawei's global competitiveness in mobile and broadband sectors, though primarily serving internal needs in a fabless model dependent on foundries like TSMC.[25] This period marked HiSilicon's transition from niche telecom supplier to broad-spectrum chip designer, underpinning Huawei's rise without external sales until later years.[2]Adaptation to Sanctions (2019-2025)
In May 2019, following the U.S. Department of Commerce's addition of Huawei to the Entity List, HiSilicon anticipated disruptions by stockpiling semiconductor components, including at least a three-month supply of critical chips to sustain production amid export restrictions on U.S.-origin technology.[26] This preemptive measure allowed HiSilicon to continue manufacturing advanced Kirin processors, such as the Kirin 990 on TSMC's 7nm node, for devices like the Huawei Mate 30 series launched in September 2019, before full supply chain severed.[27] However, escalating rules in May 2020, which required foreign foundries using U.S. equipment to obtain licenses for Huawei-bound chips, accelerated the depletion of stockpiles, with HiSilicon's global smartphone chip market share dropping to zero by Q3 2022.[28] Faced with halted access to leading foundries like TSMC and intellectual property challenges—including temporary disruptions in ARM architecture licenses—HiSilicon pivoted to domestic alternatives, partnering closely with Semiconductor Manufacturing International Corporation (SMIC) to leverage China's nascent advanced-node capabilities.[29] This shift emphasized self-reliance, aligning with Beijing's broader push for indigenous semiconductor production, though initial outputs relied on older process nodes like 14nm due to SMIC's limitations in extreme ultraviolet (EUV) lithography access.[30] By 2023, HiSilicon achieved a milestone with the Kirin 9000S processor in the Huawei Mate 60 Pro, fabricated by SMIC on its 7nm (N+2) FinFET process without EUV, marking China's first commercial-scale deployment of such technology and enabling 5G connectivity despite sanctions.[31][32] Performance benchmarks showed the Kirin 9000S trailing its pre-sanctions predecessor, the TSMC-made Kirin 9000 on 5nm, in multi-core tasks but sufficient for competitive smartphone use, underscoring adaptive engineering amid node constraints.[33] From 2023 to 2025, HiSilicon expanded this strategy to AI and server chips, producing the Ascend 910B accelerator on SMIC's 7nm process to support domestic AI training amid U.S. restrictions on Nvidia equivalents, with Huawei shipping approximately 700,000 units in 2025 to fuel China's computing infrastructure.[34] Heavy R&D investments—spurred by sanctions—drove revenue growth, with HiSilicon reportedly doubling earnings by mid-2025 through strong domestic demand for smartphones and AI hardware, though overall Huawei operations recorded quarterly losses from these expenditures.[35][36] Despite these advances, persistent gaps in sub-7nm yields and efficiency limited scalability, as SMIC's deep ultraviolet lithography-dependent processes yielded lower throughput than global leaders, highlighting sanctions' role in accelerating but constraining China's semiconductor autonomy.[37][38]Core Technologies
Arm-Based Processor Architectures
HiSilicon's Arm-based processor architectures primarily revolve around implementations of the ARM instruction set architecture (ISA), which the company licenses from Arm Holdings to design system-on-chips (SoCs) for mobile devices, servers, and other applications. These architectures integrate CPU cores compatible with ARMv8-A and later extensions, often combining performance-oriented big cores with efficiency-focused little cores in a heterogeneous configuration to optimize power and performance. Early designs relied heavily on Arm's off-the-shelf Cortex-A CPU IP, such as the Cortex-A53, A57, and A77, integrated into the Kirin mobile SoC series for balanced workloads in smartphones and tablets.[39][40] In response to U.S. export restrictions imposed in 2019, which temporarily disrupted access to advanced Arm IP, HiSilicon accelerated development of custom microarchitectures while adhering to licensed ARMv8 specifications. The Taishan series represents HiSilicon's proprietary CPU cores, first introduced for server applications in the Kunpeng lineup. The Taishan v110, debuting in the Kunpeng 920 SoC announced in 2019, features a 4-way superscalar, out-of-order execution pipeline supporting up to 64 cores per socket at frequencies reaching 2.6 GHz, fabricated initially on TSMC's 7 nm process. This custom design prioritizes multi-threaded server workloads, with each core including 1 MB of private L3 cache and support for DDR4 memory channels. Subsequent iterations, such as the Taishan v120 unveiled around 2023, achieve single-core performance comparable to AMD's Zen 3 architecture from 2020, as evidenced by benchmarks showing equivalent scores in tasks like SPECint.[41][42][43] For mobile processors, HiSilicon adapted Taishan-derived cores into the Kirin series post-sanctions, diverging from pure Cortex reliance. The Kirin 9000S, launched in 2023 for the Huawei Mate 60, incorporates four Taishan-based big cores with simultaneous multithreading (SMT), clocked up to 2.62 GHz, alongside efficiency cores, enabling competitive performance despite domestic 7 nm-class fabrication. Later models like the Kirin 9010, released in 2024, employ six medium-performance Taishan cores at up to 2.3 GHz and four Cortex-A510 efficiency cores at 1.55 GHz, supporting 5G modems and AI tasks while maintaining ARMv8 compatibility. These custom evolutions allow HiSilicon to circumvent some licensing hurdles by leveraging pre-existing Arm architectural licenses for ISA implementation, though ongoing U.S. controls have prompted exploration of alternatives like RISC-V for future diversification.[44][45][40]Da Vinci AI Architecture
The Da Vinci architecture constitutes HiSilicon's proprietary design for neural processing units (NPUs), emphasizing scalable computation for deep neural networks across mobile system-on-chips (SoCs) and data center accelerators. Introduced in 2019 with the Kirin 810 SoC, it integrates specialized hardware for tensor arithmetic, vector operations, and scalar control to handle AI workloads efficiently while minimizing power consumption.[46][47] The architecture prioritizes a balanced memory hierarchy to mitigate bandwidth bottlenecks, enabling seamless data flow between on-chip caches and external memory.[48] At its core, Da Vinci employs heterogeneous processing units: Cube units for high-throughput matrix multiplications, delivering up to 4096 FP16 multiply-accumulate operations (MACs) or 8192 INT8 MACs per core; Vector units supporting 2048-bit wide operations in INT8, FP16, and FP32 formats with built-in functions for activations and non-maximum suppression; and Scalar units for branching and control tasks.[48] This modular setup allows instantiation of core variants—such as Tiny for lightweight inference, Lite for balanced mobile AI, and Max for intensive training—scalable from single-core mobile implementations to multi-core server configurations. For instance, the Ascend 310 accelerator incorporates two Da Vinci cores, achieving 8 TFLOPS in FP16 or 16 TOPS in INT8.[49][50] Evolutions include Da Vinci 2.0, deployed in the 2020 Kirin 9000 series with dual Ascend Lite cores plus one Tiny core, enhancing support for complex tasks like real-time image synthesis via integrated ISP linkages.[51] Later iterations, as in the 2024 Kirin 9010, sustain the architecture's focus on LPDDR5X memory integration and framework compatibility for on-device AI, such as computer vision and natural language processing.[52] In server-grade Ascend chips, like the 910 series, up to 32 Da Vinci Max cores enable competitive FP16 performance against Western counterparts, produced on domestic 7nm processes amid supply constraints.[6][53] The design's emphasis on mixed-precision computing and operator fusion reduces latency, though real-world efficacy depends on software optimization via Huawei's CANN framework.[48]Product Portfolio
Mobile System-on-Chips (Kirin Series)
The Kirin series consists of ARM-based system-on-chip (SoC) designs developed by HiSilicon for mobile applications, including smartphones, tablets, and other consumer devices, with primary deployment in Huawei products. Launched in the early 2010s as a successor to the K3 series, the lineup evolved from quad-core configurations to advanced octa-core processors incorporating dedicated neural processing units (NPUs) for AI tasks and integrated 5G modems. Key milestones include the Kirin 910, HiSilicon's first quad-core SoC released in 2012, which powered early Huawei devices like the Ascend P1, and the Kirin 960 introduced in 2016 for the Huawei Mate 9, featuring a 16nm process and octa-core setup with Mali-G71 GPU for improved graphics performance.[54][55] Subsequent flagships advanced process nodes and efficiency: the Kirin 970 (2017, 10nm) integrated HiSilicon's first NPU for on-device AI acceleration, enabling features like real-time scene recognition in cameras; the Kirin 980 (2018, 7nm FinFET) marked the series' entry into sub-10nm fabrication, with a dual-NPU setup and Cat.21 LTE modem supporting up to 1.4 Gbps downloads; and the Kirin 990/990 5G (2019, 7nm EUV) added integrated 5G capabilities via the Balong 5000 modem, achieving peak downloads of 4.6 Gbps in sub-6GHz bands. The pinnacle pre-sanctions model, Kirin 9000 (2020, 5nm), featured a 1+3+4 CPU cluster using Cortex-A77 cores, a 24-core Mali-G78 GPU, and an upgraded Da Vinci NPU, powering the Huawei Mate 40 series with 15.3 billion transistors for enhanced multitasking and 8K video processing. Mid-range variants like the Kirin 710 (2018, 14nm) and Kirin 810 (2019, 7nm) targeted budget devices, offering octa-core Cortex-A76/A55 mixes with Mali-G51 GPUs for balanced power efficiency.[2][56] U.S. export controls imposed in 2019 restricted HiSilicon's access to advanced foreign foundry capacity, such as TSMC's 5nm and below, halting production of cutting-edge Kirin chips reliant on imported equipment and IP. This led to a multi-year gap in flagship releases, with Huawei relying on stockpiled Kirin 990 inventory until 2023, when the Kirin 9000S debuted in the Mate 60 series using SMIC's 7nm N+2 process; it employed custom Taishan V120 CPU cores (1x2.62 GHz + 3x2.15 GHz) alongside Cortex-A510 efficiency cores, but benchmarked lower than the original 9000 due to denser packing and yield limitations, with Geekbench scores around 1,000 single-core versus 9000's 1,200+. Subsequent models like the Kirin 9010 (2024) maintained similar 7nm architectures for incremental improvements in P-series devices. By 2025, Huawei disclosed the Kirin 9020 in Pura 80 smartphones, reportedly achieving 40% performance gains over predecessors through optimized domestic lithography, though specifics on core counts and node remain guarded amid ongoing restrictions. These adaptations underscore HiSilicon's shift to self-reliance, prioritizing functionality over parity with competitors like Qualcomm's Snapdragon series on sub-4nm nodes.[57][58][59]| Model | Release Year | Process Node | CPU Configuration | Notable Features |
|---|---|---|---|---|
| Kirin 970 | 2017 | 10nm | 4x Cortex-A73 + 4x A53 | First integrated NPU for AI |
| Kirin 980 | 2018 | 7nm | 2x Cortex-A76 + 2x A76 + 4x A55 | Dual NPU, first 7nm Kirin |
| Kirin 990 5G | 2019 | 7nm EUV | 2x Cortex-A76 + 2x A76 + 4x A55 | Integrated Balong 5000 5G modem |
| Kirin 9000 | 2020 | 5nm | 1x Cortex-A77 + 3x A77 + 4x A55 | 24-core GPU, advanced AI processing |
| Kirin 9000S | 2023 | 7nm (SMIC) | 1x Taishan V120 + 3x V120 + 4x A510 | Post-sanctions domestic production |
Integrated Modems (Balong Series)
The Balong series consists of HiSilicon's baseband modem chipsets, engineered for integration into system-on-chips (SoCs) such as the Kirin family, enabling multi-generation cellular support in Huawei's mobile devices including smartphones and tablets. These modems prioritize multimode compatibility, from 2G/3G fallback to advanced LTE and 5G, with optimizations for power efficiency and global band coverage when embedded within larger SoC architectures.[60] Early iterations focused on LTE advancements, while later models introduced 5G capabilities, often achieving industry-first benchmarks in speed and architecture support. Integration allows for reduced latency in data handoffs and unified power management between the application processor and modem cores. Development began with foundational LTE models. The Balong 700 marked the series' debut as the first chipset supporting both LTE TDD and FDD modes under 3GPP Release 8 protocols, featuring 4x2/2x2 SU-MIMO for enhanced throughput in early 4G deployments, though limited to LTE without legacy voice integration.[61] The Balong 710, released in 2012, advanced to LTE Category 4 with 2x2/4x2 MIMO, achieving 150 Mbps downlink and 50 Mbps uplink in FDD mode, alongside 2G/3G voice support and worldwide frequency bands for roaming; it was designed for pairing with SoCs like the K3V2.[62] Subsequent enhancements included the Balong 720 for 5-mode LTE Category 6 with dual-carrier aggregation over 40 MHz bandwidths, and the Balong 765, which introduced 8x8 MIMO—the only such implementation at the time—delivering up to 1.6 Gbps downlink in LTE Category 21 configurations, as integrated in variants like the Kirin 990 4G.[63][64] The transition to 5G emphasized multimode integration for backward compatibility. The Balong 5G01, announced in February 2018, became the first commercially viable 3GPP-compliant 5G chipset, supporting sub-6 GHz bands with peak downlink speeds of 2.3 Gbps, initially targeted for customer-premises equipment but adaptable for SoC embedding.[65] This paved the way for the Balong 5000, unveiled in January 2019 as the first 7 nm multimode 5G modem encompassing 2G through 5G on a single die, with standalone (SA) and non-standalone (NSA) architecture fusion, sub-6 GHz downloads up to 4.6 Gbps, and mmWave up to 6.5 Gbps. Integrated into Kirin SoCs such as the Kirin 990 5G and Kirin 9000, it enabled devices like the Huawei Mate 20 X 5G to achieve commercial 5G connectivity while maintaining 4G fallback efficiency.[66] These integrated designs reduced component count and improved thermal performance in handsets, though production scaled amid supply constraints post-2019.| Model | Key Supported Standards | Peak Downlink Speed | Notable Integration Features |
|---|---|---|---|
| Balong 710 | LTE Cat 4 FDD/TDD, 2G/3G | 150 Mbps (FDD) | Paired with early Kirin SoCs for voice/data fallback[62] |
| Balong 765 | LTE Cat 21, 8x8 MIMO | 1.6 Gbps | Embedded in Kirin 990 4G for high-end 4G optimization[64] |
| Balong 5G01 | 5G sub-6 GHz (3GPP R15) | 2.3 Gbps | Early 5G SoC adaptability with legacy modes[65] |
| Balong 5000 | 5G SA/NSA, 2G-5G multimode | 4.6 Gbps (sub-6 GHz) | 7 nm integration in Kirin 990/9000 for smartphones[66] |
Wearable and IoT Processors
HiSilicon has developed specialized system-on-chips (SoCs) for wearable devices, with the Kirin A1 serving as a flagship example tailored for low-power Bluetooth connectivity. Announced in September 2019, the Kirin A1 (also known as Hi1132) is the world's first dual-mode Bluetooth 5.1 and Bluetooth Low Energy (BLE) 5.1 wearable chip, featuring isochronous dual-channel transmission for audio, a 356 MHz dedicated audio processor, and an Arm Cortex-M7 microprocessor for efficient task handling.[67][68] It supports applications in true wireless stereo (TWS) earbuds, smartwatches, headbands, neckbands, smart speakers, and eyewear, powering devices such as the Huawei Watch GT 2 and FreeBuds 3.[69][70] These wearable SoCs emphasize ultra-low power consumption and integrated RF capabilities to enable seamless connectivity in battery-constrained environments, with HiSilicon's designs primarily targeting wrist-worn devices like smart bands and watches, as well as head-mounted audio products.[71] In response to production challenges from sanctions, Huawei has increased output of Kirin-series chips for smart wearables as of June 2024, aiming to bolster ecosystem expansion amid restricted access to advanced nodes.[72] For Internet of Things (IoT) applications, HiSilicon provides connectivity-focused processors emphasizing wide-area coverage, low power, and compatibility with standards like NB-IoT and LTE Cat-1. Key offerings include the Hi2115, a 3GPP Release 14-compliant Cat-NB2 chip launched for ultra-low-power scenarios, integrating an MCU, memories, and interfaces while operating in bands from 698-960 MHz and 1695-2180 MHz with three Arm Cortex-M0+ cores.[73][74] The Hi2120 complements this as another cellular IoT chipset for reliable performance in smart homes, mobility, and urban infrastructure.[74] Additionally, the Hi2131, a low-power Cat-1 chip introduced in 2025, targets broader IoT deployment by enhancing connectivity efficiency.[75] HiSilicon's IoT processors also incorporate RISC-V architectures for versatility, as seen in the Hi3861 WLAN SoC for 2.4 GHz connectivity in compact development boards measuring 2 cm x 5 cm, integrating 802.11b/g/n baseband and RF circuits.[76] These chips power diverse devices including IP cameras and surveillance systems, with HiSilicon components embedded in tens of millions of global IoT units as of 2018, despite geopolitical scrutiny over supply chain integrations.[77] Recent initiatives by HiSilicon's Shanghai arm, as of July 2025, include RISC-V-based solutions for industrial IoT, smart homes, and visual AI, reflecting efforts to diversify beyond Arm dependencies.[78][79]Server Processors (Kunpeng Series)
The Kunpeng series comprises Arm-based server processors developed by HiSilicon for high-performance computing in data centers, cloud infrastructure, and enterprise servers.[80] These CPUs leverage the Armv8 architecture to deliver multi-core scalability, targeting workloads such as virtualization, databases, and big data processing.[81] Initial development of the series, internally designated Hi16xx, began around 2015, with early models produced on 16 nm processes featuring up to 32 cores but limited cache efficiency compared to contemporary x86 alternatives.[82] The flagship Kunpeng 920, introduced in January 2019, represents a significant advancement, fabricated on a 7 nm process node and incorporating HiSilicon's custom TaiShan V110 cores.[83] This 64-core configuration operates at up to 2.6 GHz base frequency (with select variants reaching 3.0 GHz), supports 8-channel DDR4 memory, PCIe 4.0 interfaces, and coherent interconnects for up to 4-socket systems with 480 Gbit/s inter-chip bandwidth.[80] The TaiShan V110 is a 64-bit Armv8-A compliant core with 4-wide out-of-order execution, marking HiSilicon's debut in proprietary core design rather than relying solely on licensed Arm IP.[41] Variants like the Kunpeng 920-6426 scale to 64 cores per socket, while lower-core options (e.g., 48-core or 8-core desktop derivatives) enable broader deployment in Huawei's TaiShan server lineup.[84] Performance evaluations indicate the Kunpeng 920 excels in core-count-heavy scenarios due to its density, outperforming Intel's Xeon Gold 6248 in multi-threaded SPEC benchmarks thanks to superior parallelism, though it lags behind AMD's Epyc processors by nearly a factor of two in similar tests. Single-threaded efficiency remains a relative weakness compared to x86 incumbents, reflecting Arm's historical trade-offs in instruction-level parallelism and branch prediction.[41] Adoption has been concentrated in domestic Chinese markets, powering Huawei's TaiShan 2280 V2 servers for applications in government, finance, and telecom sectors, with ecosystem partners certifying software compatibility for Linux distributions and open-source stacks.[82] Subsequent iterations, such as the Kunpeng 916 series, offer refined power envelopes (e.g., 64 cores at lower TDP for energy-sensitive deployments), but broader global penetration has been constrained by U.S. export restrictions limiting access to advanced fabrication.[85]AI Accelerators (Ascend Series)
The Ascend series represents HiSilicon's dedicated lineup of AI accelerators, optimized for deep learning training, inference, and hybrid workloads across edge, cloud, and data center deployments. Introduced in 2019, the series leverages Huawei's proprietary Da Vinci architecture, which features specialized AI cores for matrix-vector operations, scalar processing, and efficient data flow tailored to neural network computations.[48] This architecture enables scalable performance by integrating multiple Da Vinci cores with on-chip buffers, network-on-chip (NoC) interconnects, and support for formats like FP16 and INT8, prioritizing energy efficiency over general-purpose computing.[49] Despite U.S. export controls imposed since 2019 restricting access to advanced foundry nodes and EDA tools, HiSilicon has advanced the series through domestic fabrication at SMIC and strategic stockpiling, achieving iterative improvements in yield and integration.[86][87] The entry-level Ascend 310, launched in 2019 for edge applications such as video analytics and object recognition, integrates two Da Vinci Max AI cores alongside ARM Cortex-A55 CPU cores, delivering 16 TOPS at INT8 precision with only 8 W power consumption.[88] It supports 16-channel H.264/H.265 video decoding and is deployed in compact modules like the Atlas 200 for on-device AI processing in industrial and IoT scenarios.[89] Higher-end models target data center training; the flagship Ascend 910, unveiled on August 23, 2019, was positioned as the world's most powerful AI processor at launch, with 32 Da Vinci cores in four clusters, 1024-bit NoC at 2 GHz, and interfaces for high-bandwidth memory.[90] Subsequent iterations include the Ascend 910B (2022), comparable to Nvidia's A100 in select benchmarks, and the Ascend 910C (mass production starting May 2025), which fuses two 910B dies to achieve approximately 60% of Nvidia H100 performance in AI training tasks while optimizing for China's domestic ecosystem.[91][92]| Model | Da Vinci Cores | Key Performance (INT8/FP16) | Power | Target Use Case | Release Year |
|---|---|---|---|---|---|
| Ascend 310 | 2 | 16 TOPS / 8 TFLOPS | 8 W | Edge inference | 2019 |
| Ascend 910 | 32 | 512 TOPS / 256 TFLOPS | 350 W | Data center training | 2019 |
| Ascend 910C | Dual 910B | ~60% H100 equiv. | N/A | Large-scale clusters | 2025 |
Manufacturing and Supply Chain
Global Foundry Dependencies
HiSilicon, as a fabless semiconductor design firm, relies entirely on external foundries for wafer fabrication, with a heavy dependence on global leaders like TSMC for cutting-edge processes unavailable domestically prior to geopolitical restrictions.[99] From its inception, HiSilicon outsourced production of high-volume products such as the Kirin series mobile SoCs and Ascend AI accelerators to TSMC, which manufactured chips on 7nm and earlier nodes; for instance, the Kirin 990, introduced in 2019, was fabricated by TSMC on its 7nm process.[100] This partnership peaked in 2019, when Huawei and HiSilicon accounted for approximately 37% of TSMC's total sales, underscoring the foundry's critical role in enabling HiSilicon's competitive edge in smartphones and AI hardware.[99] The U.S. Department of Commerce's expansion of the Foreign Direct Product Rule (FDPR) in August 2020 effectively severed this supply chain, prohibiting foreign foundries using U.S.-origin technology—such as TSMC's processes reliant on American equipment and IP—from producing for Huawei entities without licenses.[101] TSMC halted new orders from HiSilicon on September 14, 2020, though pre-existing inventory allowed limited continuity for products like the Kirin 9000, also on TSMC's 5nm node.[102] This cutoff exposed HiSilicon's vulnerability to international supply disruptions, as alternative global foundries like Samsung were either unwilling or unable to fill the gap due to similar U.S. extraterritorial controls, forcing a pivot toward less advanced domestic alternatives.[27] Post-2020, attempts to circumvent restrictions, such as alleged proxy arrangements uncovered by TSMC in 2024 involving precursor production for Ascend chips, highlight ongoing tensions but have not restored full access to global advanced nodes.[103] HiSilicon's pre-sanctions reliance on TSMC not only accelerated its technological roadmap but also amplified risks from geopolitical dependencies, with no evidence of diversified global partnerships mitigating the 2020 impact.[104]Domestic Production Advances
In response to U.S. export restrictions imposed in 2020, which barred access to advanced foreign foundries like TSMC, HiSilicon pivoted to domestic manufacturing partnerships, primarily with Semiconductor Manufacturing International Corporation (SMIC). This shift enabled the production of the Kirin 9000S system-on-chip in 2023, fabricated entirely on SMIC's 7 nm N+2 process for the Huawei Mate 60 Pro smartphone, marking a significant milestone in China's advanced node capabilities without extreme ultraviolet (EUV) lithography.[105][106] The 7 nm N+2 node employs deep ultraviolet (DUV) lithography with multi-patterning techniques to achieve feature densities comparable to some global standards, though at elevated costs—up to 50% premiums over TSMC equivalents—and with larger critical dimensions that limit performance efficiency.[105][107] By mid-2025, HiSilicon continued leveraging this process for designs like the Kirin X90 in the Huawei Matebook Fold, underscoring persistent challenges in scaling to 5 nm production despite developmental progress at SMIC.[108][109] In the AI domain, HiSilicon's Ascend 910C accelerator advanced toward mass production in the first quarter of 2025 via SMIC's 7 nm capacities, bolstering domestic high-performance computing amid broader national self-reliance initiatives.[110][111] SMIC's reported trajectory includes 5 nm node completion by late 2025, potentially enabling future HiSilicon iterations, though yields remain constrained and overall self-sufficiency in cutting-edge semiconductors trails global leaders, with China achieving only partial progress toward its 70% domestic production target under the Made in China 2025 plan.[107][112][113]Geopolitical Challenges
US Export Controls and Entity List
In May 2019, the United States Department of Commerce's Bureau of Industry and Security (BIS) added HiSilicon Technologies Co., Ltd., headquartered in Shenzhen, China, to the Entity List under the Export Administration Regulations (EAR), effective May 16, 2019, as part of a broader action targeting Huawei Technologies Co., Ltd. and 68 affiliated entities.[114] This designation was based on BIS's determination that HiSilicon, Huawei's semiconductor design subsidiary, was involved in activities contrary to U.S. national security or foreign policy interests, including risks associated with the development and supply of technologies potentially enabling surveillance or military applications.[115] Placement on the Entity List imposes a licensing requirement for any export, reexport, or transfer (in-country) of items subject to the EAR—including commodities, software, and technology—to HiSilicon, with license applications subject to a presumption of denial to restrict access to U.S.-origin goods and technology. The Entity List action disrupted HiSilicon's supply chain, particularly its reliance on U.S.-controlled intellectual property and equipment for chip design and fabrication processes. In August 2019, BIS added 46 additional Huawei non-U.S. affiliates to the list, further encompassing HiSilicon's operational network.[116] On May 15, 2020, BIS amended the foreign direct product rule (FDPR) with a "Huawei-specific" expansion under §734.9(e)(1) of the EAR, requiring licenses for foreign-produced items that are direct products of certain U.S. technologies, software, or equipment if destined for HiSilicon or other Huawei footnote 1 entities, even if manufactured abroad.[117] This rule change, effective immediately with a 120-day grace period for some compliance, targeted circumvention risks and severely limited HiSilicon's access to advanced semiconductor manufacturing, as key foundries like TSMC incorporate U.S.-origin tools and IP subject to the expanded FDPR.[118] Subsequent BIS actions have reinforced these controls, including August 2020 additions of 38 more Huawei affiliates and ongoing enforcement, such as a November 2021 settlement fining a U.S. firm $80,000 for unlicensed exports of controlled equipment to HiSilicon.[119][120] The Entity List status remains in effect as of 2025, with HiSilicon designated under a "footnote 1" annotation that triggers the FDPR, prohibiting unlicensed supply of specified semiconductors and related items produced with U.S. technology. These measures have compelled HiSilicon to pivot toward domestic alternatives, though persistent U.S. technology dependencies in global semiconductor ecosystems continue to constrain its output of high-performance chips like the Kirin and Ascend series.[121]Alleged Security Concerns and Denials
In February 2020, Russian security researcher Vladislav Yarmak disclosed a backdoor vulnerability in firmware for digital video recorders (DVRs) and network video recorders (NVRs) using HiSilicon system-on-chips (SoCs), such as the Hi3520 series, where hardcoded credentials allowed remote root access without authentication.[122] [123] The issue affected devices from third-party manufacturers like Hangzhou Xiongmai Technology, enabling potential unauthorized surveillance or control, though Yarmak attributed it to firmware implementation rather than the HiSilicon hardware itself.[124] HiSilicon's analysis concluded the vulnerability originated in customer-supplied software stacks, not the chips or provided SDKs, and emphasized that no deliberate backdoors were embedded by the company.[125] [126] Broader allegations against HiSilicon stem from its parent company Huawei's designation as a national security risk by the U.S. government, which in 2019 added Huawei (including affiliates like HiSilicon) to the Entity List, citing risks of espionage and intellectual property theft under China's National Intelligence Law, which mandates corporate cooperation with state intelligence efforts.[127] U.S. officials have expressed concerns that HiSilicon-designed chips in telecommunications and surveillance equipment could facilitate undisclosed access for Chinese authorities, potentially disrupting critical infrastructure or enabling data exfiltration, though no public forensic evidence has confirmed hardware-level backdoors in HiSilicon products.[128] Independent assessments, including those from security firms, have identified software vulnerabilities in HiSilicon-based devices but found no systematic evidence of state-sponsored espionage mechanisms embedded in the chips.[129] HiSilicon and Huawei have consistently denied introducing backdoors or vulnerabilities for unauthorized access, stating in official notices that the company adheres to international cybersecurity standards and conducts rigorous internal reviews, with any reported issues traced to third-party integrations rather than core chip design.[125] [130] In response to the 2020 DVR backdoor claims, HiSilicon urged customers to update firmware and collaborated with vendors for patches, while rejecting assertions of intentional flaws.[131] Critics, including U.S. policymakers, maintain skepticism due to opaque supply chains and legal obligations under Chinese law, but empirical demonstrations of exploitable backdoors remain limited to post-manufacture firmware modifications, not inherent chip architecture.[132]Industry Workarounds and Responses
In response to U.S. export controls imposed after Huawei's addition to the Entity List in May 2019, HiSilicon pivoted to domestic manufacturing partners, primarily Semiconductor Manufacturing International Corporation (SMIC), to sustain chip production. This shift enabled the fabrication of system-on-chips like the Kirin 9000S, a 7nm-class processor released in August 2023 for the Huawei Mate 60 smartphone, using SMIC's N+2 process node derived from deep ultraviolet (DUV) lithography with multi-patterning techniques rather than prohibited extreme ultraviolet (EUV) equipment.[133][33] Despite lower yields and performance deficits compared to TSMC's 5nm Kirin 9000—evidenced by benchmark tests showing reduced CPU and GPU efficiency—the Kirin 9000S demonstrated circumvention of restrictions barring access to advanced foreign tools.[33][134] For AI accelerators, HiSilicon produced the Ascend 910B using SMIC's 7nm process, incorporating domestic alternatives to restricted U.S.-origin intellectual property while relying on architectural optimizations such as chiplet designs and software stacking to approximate higher performance.[37] Huawei CEO Ren Zhengfei acknowledged in June 2025 that the firm trails U.S. technology by one generation but compensates through methods like cluster computing and enhanced algorithms, prioritizing volume production for domestic markets capped at around 200,000 Ascend units annually due to ongoing controls.[135][136] These efforts align with China's broader semiconductor self-reliance strategy, including state-backed investments exceeding $150 billion since 2014, which have accelerated SMIC's progress toward 5nm nodes using DUV-based multi-patterning for transistor densities up to 130 million per square millimeter.[137][138] Industry-wide, Chinese entities have responded by verticalizing supply chains under Huawei's influence, with HiSilicon integrating design, testing, and packaging to minimize foreign dependencies, as seen in its dominance of domestic fabless production.[139] However, U.S. authorities have countered with expanded rules, such as the 2020 foreign direct product rule amendments, targeting foreign-made items incorporating U.S. technology, prompting further adaptations like indigenous EDA tools and materials sourcing.[140] Yields remain a bottleneck, with SMIC's 7nm processes reportedly achieving only 30-40% efficiency versus global leaders, underscoring the controls' partial efficacy in constraining advanced scaling.[38]Market Impact and Performance
Revenue Trends and Market Share
HiSilicon's revenue experienced significant volatility following the U.S. Department of Commerce's addition of Huawei and its affiliates to the Entity List in May 2019, which restricted access to advanced manufacturing processes and U.S. technology. Prior to these restrictions, HiSilicon contributed substantially to Huawei's device ecosystem, but by Q3 2022, its global smartphone chipset market share had declined to 0% due to inability to produce competitive advanced nodes.[141] A resurgence began in 2023-2024, driven by domestic advancements in chip design and production at nodes like 7nm via SMIC, enabling Huawei's return to premium smartphone segments with Kirin processors. HiSilicon's revenue doubled year-over-year in 2024, fueled by strong sales of Huawei's Pura 70 and Mate 70 series handsets, which incorporated in-house SoCs.[142][35] This growth aligned with Huawei's consumer business revenue surging 38% year-on-year to 339 billion yuan (US$47 billion) in 2024, reflecting increased internal procurement of HiSilicon chips amid broader Huawei revenues reaching 862 billion yuan, nearing the 2020 pre-sanctions peak.[143] In market share terms, HiSilicon captured 12% of the global premium Android smartphone system-on-a-chip (SoC) market in 2024, up from 8% in 2023, ranking third in global smartphone chip revenues for that year.[142][144] Within China, HiSilicon has solidified dominance in Huawei's supply chain, supporting the company's 18.1% smartphone market share as of 2025, though broader semiconductor market shares remain constrained by export controls limiting external sales, with over 90% of output directed internally to Huawei.[139] Projections for 2025 indicate continued internal revenue growth tied to Huawei's anticipated exceedance of 900 billion yuan overall revenue, bolstered by 5G-enabled devices, though high R&D spending—reaching 22.7% of Huawei's H1 2025 sales—may pressure net profits amid geopolitical barriers to global expansion.[145][146]Technological Contributions and Benchmarks
HiSilicon's Kirin series of mobile processors represents a major contribution to integrated SoC design, incorporating custom Taishan CPU cores derived from ARMv8 architecture, Mali GPUs, and dedicated Neural Processing Units (NPUs) for on-device AI tasks. The Kirin 9020, featured in Huawei's Mate 70 series launched in late 2024, utilizes a 7nm process node fabricated by SMIC, with an octa-core configuration including high-performance Taishan V120 cores clocked up to 2.5 GHz, delivering AnTuTu v10 scores exceeding 1 million points and Geekbench 6 single-core results around 1,500, positioning it competitively against Qualcomm's Snapdragon 8 Gen 3 in multi-threaded workloads. Earlier iterations like the Kirin 980, introduced in 2018 on TSMC's 7nm node, pioneered dual-NPU setups for enhanced machine learning efficiency, achieving Geekbench 5 multi-core scores of approximately 3,000 in devices such as the Huawei P30. These designs emphasize power efficiency and 5G integration via Balong modems, with the Kirin 9000 (2020) supporting sub-6GHz and mmWave bands while integrating ISP for advanced imaging. In AI acceleration, HiSilicon's Ascend series advances data center-scale computing through the Da Vinci architecture, which optimizes tensor operations and supports large-scale clustering. The Ascend 910C, released in 2025, offers up to 800 TFLOPS in FP16 precision with 128 GB HBM3 memory at 310W TDP, enabling Huawei's CloudMatrix 384 cluster to surpass Nvidia's H800 in inference benchmarks for models like DeepSeek R1, reportedly achieving higher tokens-per-second throughput in reasoning tasks due to optimized interconnects and software stack integration. Independent tests indicate the Ascend 910B predecessor delivers around 60% of Nvidia H100's performance in mixed-precision AI training, bolstered by innovations in multi-chip module packaging to mitigate yield issues on SMIC's 7nm node. These chips contribute to China's push for AI sovereignty, with Huawei claiming up to 80 TFLOPS FP16 uplift per generation through architectural refinements like enhanced matrix multiply units. HiSilicon's Kunpeng and Taishan server processors extend ARM-based computing to high-performance data centers, featuring up to 64 cores per die on 7nm processes with coherent interconnects for multi-socket scalability. The Kunpeng 920, announced in 2019, operates at 2.6 GHz across 64 cores, yielding SPEC CPU 2017 integer rate scores of around 930 in base configuration and supporting 8-channel DDR4 memory for bandwidth-intensive workloads. The Taishan V120 core, refined in later iterations, matches AMD Zen 3 single-core performance in leaked benchmarks from 2024, with out-of-order execution widths comparable to contemporary x86 designs, enabling efficient virtualization and cloud-native applications in Huawei's TaiShan servers. Contributions include heterogeneous integration with openEuler OS for embedded deployments and early adoption of ARMv8.2 extensions for cryptography acceleration.| Processor Series | Key Innovation | Representative Benchmark |
|---|---|---|
| Kirin (Mobile SoC) | Custom Taishan cores + integrated 5G/NPU | Kirin 9020: AnTuTu v10 >1M; Geekbench 6 single ~1,500[147] |
| Ascend (AI Accelerator) | Da Vinci tensor core + cluster scaling | 910C CloudMatrix: Outperforms H800 in DeepSeek R1 inference[148] |
| Kunpeng/Taishan (Server) | 64-core ARMv8 + multi-socket coherence | Kunpeng 920: SPECint_base 930 @ 2.6 GHz[149] |