Research & Innovation
The intelligent world is approaching, faster than ever before. This is creating unprecedented opportunities and challenges in terms of innovation. Huawei continues to invest in basic research and open innovation, and accommodates and addresses customer needs with an open mind, while steering these needs with science and technology. We have also built flexible business models and encourage numerous industries to adopt a vast range of models and applications. Through such initiatives, we are bringing digital to every person, home and organization for a fully connected, intelligent world.
Basic Research
In 2023, we continued to focus on basic research and strengthened efforts to steer customer needs with cutting-edge technologies. We also pursued breakthroughs in innovation that will drive the entire industry and all of society forward.
When exploring the infinite possibilities of science, we remain dedicated to combining mathematics with modern information and communications technologies. In 2023, we further developed our basic research capabilities and broke through numerous bottlenecks plaguing the industry:
- In information theory coding, we extended the classical fast Fourier transform (FFT) algorithm to algebraic geometry codes by Goppa, reducing their complexity from O(n^2) to O(nlog n).
- In machine learning, we developed a new federated conformal prediction method based on quantile regression. This method addresses the label shift between agents and provides theoretical guarantees for both valid coverage of the prediction sets and differential privacy.
- In large language model training, we developed CAME, a novel memory-efficient optimizer that supports adaptive confidence-based updating and consumes over 50% less memory than the Adam optimizer. The paper we published on this optimizer was named one of ACL's (Association for Computational Linguistics) 2023 Outstanding Papers.
We also ramped up efforts to apply our latest basic research results to industry. At Huawei, we refer to this as "laying eggs along the way".
- In wireless communications, we developed an energy-efficient low-order digital predistortion architecture to manage nonlinear power amplifier distortion in wideband massive multiple-input multiple-output (MIMO) systems. This design is based on a decomposition theory and can realize a 3–5 dB improvement in correction performance.
- In optical communications:
- We designed a simplified nonlinear fiber compensation algorithm, using the perturbation theory to solve nonlinear Schrodinger equations. This algorithm has helped extend the long-haul transmission distance of 800 gigabit fiber by about 20%.
- We also designed a metasurface-enabled light modulation model based on the finite-difference time-domain (FDTD) method to address polarization dependence of liquid crystal on silicon (LCoS) – the engine for optical cross-connect (OXC). This model has allowed us to simplify wavelength selective switch (WSS) by anywhere from 50% to 75%. - In artificial intelligence, we introduced Integral Neural Networks, a new family of networks featuring continuous layer representation to address the intense computing resource demands of generative model training. These networks support multi-process inference on a single XPU, multiplying inference efficiency without losing accuracy.
- In information retrieval, we proposed a new method for approximating data distribution through function mapping to handle possible explosions in index space triggered by rapid increases in high-dimensional data. This method can boost retrieval speed by an order of magnitude.
Open Innovation
We have continued pursuing open innovation along multiple paths in multiple waves, targeting multiple scenarios. The company works tirelessly to unleash the full value of its complete portfolio, address real business challenges, and create value for the industry.
Wireless Communications
- We field tested prototypes in 11 scenarios and managed to:
- Become the first to verify zero-overhead integrated sensing and communication capabilities for 6G networks on the centimeter wave (cmWave).
- Realize centimeter-level precision of multi-object detection and decimeter-level reconstruction of outdoor environments on cmWave band.
- Verify polarization multiplexing and beam spatial division multiplexing technologies under ultra-large bandwidth on the terahertz frequency band, reaching a peak speed of over 400 Gbit/s at an outdoor distance of 500 meters. - We were also the first to achieve 0.5 gigabit satellite communications based on cellular mobile broadband protocols while moving at speeds of 130 km/h, demonstrating the potential of satellite-based cellular mobile network enhancement.
Optical Networks
We continued to increase single-wavelength rate and channel capacity, allowing us to:
- Make breakthroughs in 222 gigabaud ultra-high-speed optoelectronic modulator technology and propose a new key subcarrier algorithm and architecture to address coding modulation and power consumption in large throughput scenarios.
- Break new ground in the design and production of a new type of high density multi-element doped optical fiber.
- Commercially use erbium-doped fiber amplifiers (EDFAs) on ultra-wide gain spectrum: C6T + L6T (12 THz).
- Test and verify the first ever 400 gigabit quadrature phase shift keying (QPSK) system, which was able to deliver a capacity of 32T at a distance of 7,000 kilometers.
Carrier and Enterprise Networks
Our innovation in this domain in 2023 centered on three areas:
- We created a layered heterogeneous topology network architecture and a full-path routing technology system that shorten the interconnection distance between nodes and reduce the use of optical components. Therefore, network costs now account for less than 20% of total costs and network reliability has been improved by an order of magnitude.
- We invented a global routing planning algorithm to increase network throughput to over 90%.
- We created the world's first non-uniform Bruck collective communications algorithm, which delivers 50% better performance than other mainstream algorithms in the industry.
AI Algorithms
Foundation models are driving us ever closer to a truly intelligent world, but they also present a number of new challenges. Our innovations in AI algorithms focus on transforming these challenges into opportunities.
- To date, Ascend clusters have been used to train trillions of high-quality parameters for Pangu foundation models. This has helped us:
- Realize near-linear scaling of training for 4,000 XPUs and address the technological challenges impeding training and inference acceleration and long sequences;
- Boost the 7-day retention rate of Celia users on Mate 60 phones by more than 10 percentage points; and
- Use AI in Huawei Cloud EI and business process & IT services for applications in a dozen industries. - The company also released simplified network architectures that best support Ascend, such as GhostNet v2 and VanillaNet, boosting efficiency by at least 50% while performing the same task to the same level of precision.
Computing
The demand for extreme-scale computing has become a core engine of AI development, so we innovated to achieve a number of breakthroughs in this area in 2023:
- We implemented a novel supernode architecture to realize unified memory addressing and resource scheduling.
- Our new liquid-cooling architecture pushed engineering to new heights in order to reduce footprint while increasing computing power for higher-density computing systems.
- We provide an open and easy-to-use framework for computing ecosystems and have released CANN 7.0, which enables developers to customize high-performance operators and acceleration libraries that directly schedule and manage computing resources.
- We upgraded the Ascend C programming language to simplify operator implementation logic, reducing operator customization time from weeks to days and accelerating AI model and application development.
Consumer Business
Throughout 2023, Huawei remained focused on deepening our roots in core technologies to provide users with industry-leading experience.
- We pioneered advanced imaging technologies such as the adjustable aperture and Super TeleMacro Camera on our phones. We also made groundbreaking efforts in related areas such as an ultra-high dynamic range in scenes with higher light-to-dark ratios, delicate bokeh effects, and multi-modal deblurring. These incredible XMAGE technologies have brought ultra vision photography experiences to our users.
- We became one of the world's first companies to launch a cloud enhancement function backed by technologies like on-cloud algorithms for foundation models trained with hundreds of millions of parameters, a hybrid network supporting both photo preview and sensing, and high image compression. It helps phones capture and recover details within images and deliver portraits of unparalleled clarity and fineness, pushing photography to new levels of precision and beauty.
- Huawei phones now automatically connect to the most optimal network they can find, providing better communications quality and throughput in weak-signal scenarios and delivering more stable network connections.
- We unveiled our next-gen, double-rotating, and multi-dimensional hinge that is able to deliver a free range of motion thanks to its dual-track driver structure and ultra-strong aluminum. This makes HUAWEI Mate X5 phones both incredibly light and exceptionally durable.
- We continued to work on materials and production methods for our phones, developing a number of new technologies, such as Kunlun Glass, whose upgraded formula and atomic structure enhancement make it 100% more drop resistant and 300% more scratch resistant.
- Our HongMeng Kernel was awarded the industry's first Evaluation Assurance Level 6 Augmented (EAL6+) certificate as part of the Common Criteria for Information Technology Security Evaluation (CC), making Huawei the world's first smart device manufacturer to receive the certification in the commercial OS kernel field.
- The number of devices running on HarmonyOS rose to 800 million in 2023 thanks to the system's adoption across numerous industries.
- Following the revamped AITO M7's launch, over 120,000 vehicles were sold in just 100 days. This achievement can be attributed to Huawei's three major proprietary technologies – General Obstacle Detection (GOD), Road Cognition & Reasoning (RCR), and decision making under uncertainty.
Foundational Software
In 2023, we continued pursuing innovation in foundational technologies in order to create a more solid foundation for the industry and ecosystems, resulting in numerous breakthroughs, including:
- Our proprietary GaussDB is becoming China's first database to support full-stack hardware-software collaboration to enable strong dual-cluster consistency, high elasticity for thousands of nodes, and full encryption.
- openEuler now comes equipped with unique new strengths to better address industrial latency, security, and resource utilization requirements, thanks to advances in deterministic low latency, confidential computing, and hybrid deployment and scheduling.
Software Engineering
Vulnerabilities from open source software are a major source of security threats for the software supply chain. Considering this, Huawei has developed an open source vulnerability knowledge library, and made significant headway into technologies like characteristic abstraction, analysis, and comparison, as well as vulnerability research and judgment. This has helped us detect and identify vulnerabilities originating from open source C, C++, and Java projects in advance.
Systems Engineering
In L3 and higher-level complicated autonomous driving scenarios, ensuring safety of the intended functionality (SOTIF) is a great challenge. To surmount this challenge, we:
- Built a simulation testing capability that covers both hazardous scenarios and fault injection, boosting end-to-end testing efficiency by at least five times.
- Achieved leapfrog development in the out of operational design domain (out-of-ODD) detection technology, with the accuracy of safe lane planning at intersections hitting 92%.