Research & Innovation

Today, we live in a dynamic and turbulent world, where changes occur at breathtaking speed. Challenges loom large, yet the opportunities before us are unparalleled. At Huawei, we have two key drivers of innovation: science and technology, and customer needs. Both commercial value and market demands are driving our innovation and determining how we invest in science and technology. Breakthroughs in technology, in return, stimulate customer needs and allow us to create greater value for customers.

Basic Research

In 2024, we strengthened our basic research and theoretical study, aligning with business strategies and centering on mathematical logic. This has helped the company make the soil more fertile and build up innovation capabilities to break through bottlenecks holding back the development of specific businesses.

Huawei insists on basic research, as each breakthrough is like a beacon in the darkness, which will generate innovation momentum and foster long-term growth across diverse business sectors.

Game theory:

• We proposed a new prophet inequality to address the integrality gap between an infinite-dimensional mixed-integer program and its linear-programming relaxation. Regarding the long-standing problem of revenue maximization in the unit-demand single-buyer setting, our new approach guarantees a tight 3-approximation to the established theoretical benchmark, marking the first major breakthrough in this area in 15 years. The paper on this breakthrough was published at the 2024 IEEE 65th Annual Symposium on Foundations of Computer Science (FOCS).

Basic AI theories:

• We presented the first reinforcement learning and value-based tree-search learning framework for the purpose of fine-tuning open source large language models (LLMs) and improving their chain-of-thought (CoT) reasoning. This approach defines a value function for tree nodes, thereby boosting the inference accuracy of certain mathematical and gaming tasks by 14% on average and by 50% at maximum.
• We also put forward LEGO-Prover, a theorem proving agent based on LLMs. Through interaction and coordination between its prover, evolver, and growing skill library, our work facilitates automatic theorem proving, resulting in a 5%–10% improvement in success rates on miniF2F.

Our scientific research aims to bring together the ingenuity of scientists to solve the requirements, challenges, and pain points of different businesses.

• Massive MIMO transceivers feature high computational complexity. To address this, we proposed a theory and framework for transform-domain signal processing, which effectively suppresses the expansion of signals' dynamic range and reduces computational complexity by more than 25%.
• Based on the Hebbian learning rule, we developed a new local strategy, other than backpropagation, for neural networks, achieving a fine-tuning throughput improvement of over 20% for LLMs.
• We have used an algebraic modular approach and short Reed-Solomon (RS) codes to realize rapid decoding of RS codes. This has helped us approach the theoretical minimum number of iterations inherent in the Welch-Berlekamp (WB) key-equation problem, thus surpassing the t+e boundary established by C.-L. Chen in 1981.

Our Chaspark website provides online access to global academic and technical resources, connecting the world and creating fertile ground on which we can nurture new ideas through academic exchanges. This platform provides free online access to over 190 million patent records and more than 8,000 academic conferences, and has published more than 570 industry challenges, 300 contest problems, and 10 open source projects. Many university educators submit ideas to help find solutions to these challenges, and their efforts have helped resolve more than 300 such challenges to date.

Open Innovation

We dive deep into customer scenarios, stay close to end users, creatively address issues, and make significant technological breakthroughs based on our solid theoretical foundation. Guided by commercial value and market demands, we leverage our strengths in systems engineering and a complete portfolio to constantly create value for customers.

Wireless Communications

We are innovating nonstop to drive industry breakthroughs and seize enormous opportunities presented by the age of mobile AI. To this end, last year, we:

• Put forward the A-Core network architecture prototype based on agentic AI, making application-driven generative AI networks a reality.
• Broke through the pilot resource utilization bottleneck that hindered traditional massive MIMO in a centimeter wave frequency band by employing AI-based RF-MAP and full-rank calculation in ultra-low-dimensional latent space. This allowed us to increase spectrum efficiency by almost five times.
• Utilized the effects of self-healing, a property of orbital-angular-momentum, in a millimeter wave frequency band to realize non-line-ofsight propagation of signals on the ground, expanding coverage by more than 10 dB.
• Enabled direct connection of mobile phones to Low Earth Orbit satellites over a millimeter wave frequency band, increasing capacity density by around five times.

Optical Networks

In optical networking, we:

• Defined a new optical fiber channel damage model and innovatively used cascaded forward error correction (FEC) technology to realize low latency, and helped formulate the IEEE standard for short-range 800G optical interconnection.
• Made advances in large-bandwidth high-order modulation and in 240 GBd ultra-low-drive-voltage thin film lithium niobate (TFLN) modulation technology, helping build a data center interconnect (DCI) solution with the lowest possible cost per bit.
• Overcame technical challenges such as multilayer wafer hybrid bonding and wafer-level optical window hermetic packaging, allowing for the large-scale commercial use of high-port-count MEMS-based OXC. This enables the efficient scheduling of intelligent computing clusters with millions of cards.

Networking

For networking, we:

• Proposed a novel network spatio-temporal balance algorithm to overcome the challenge of sparse model training regarding parallel communications, thus realizing network throughput greater than 95%.
• Developed the industry's first collective communication algorithm for asymmetric, heterogeneous links, solving the issue of reusing bandwidth for buses and networks while improving communications performance by more than 55%.

AI Algorithms

We have made constant breakthroughs in AI algorithms, which will help us tackle key challenges and seize major opportunities related to foundation models in the age of AI.

• To date, Pangu models have been used in over 400 scenarios across more than 30 industries, helping these industries go intelligent more easily. – Pangu models have helped Celia perform training with trillions of tokens and aided its memory and perception in 23 top scenarios, translating to a task planning success rate of 90%.
• HUAWEI ADS 3.0 now comes packed with a multi-modal fusion perception network and an innovative end-to-end cognitive architecture that supports brain-like thinking. These technologies help realize lossless information transmission and end-to-end, human-like intelligent driving, with responsiveness increased by at least 100%. In 2024, HUAWEI ADS-powered vehicles drove over 1.4 billion kilometers with intelligent driving – 30% of their total mileage. The greatest distance traveled by a single user using Huawei's intelligent driving system exceeded 100,000 km. Users also used Huawei's intelligent driving system to park over 100 million times last year.

Computing

In computing, we:

• Released the AI-powered Atlas 900 A3 Cluster that comes with a new architecture that deeply integrates compute, storage, and networks. This combined with SuperPOD has helped us craft systematic computing solutions, addressing customer computing demands for model training and inference.
• Continued sharpening the competitive edge of our Ascend products and solutions by improving their performance, ease of use, and openness. – Launched CANN 8.0, which supports more than 80 new fused operators and over 100 Ascend C APIs.
• Developed a collective communication algorithm, designed the LLM PD-disaggregated deployment solution, unveiled the LLMDataDist component, and started supporting a PyTorch library for extending graph mode capabilities to enhance LLM inference performance on Ascend.

Consumer Business

We remain committed to technological innovation and foundational technology breakthroughs. We made a number of breakthroughs in this area to ensure we can consistently provide consumers with a better experience.

• We developed the industry's first Advanced Precision Hinge System for foldable phones. This system ensures precise and delicate dualhinge movement so that each fold, whether inward and outward, is smooth and seamless.
• We made great progress in ultra-thin, flexible glass technology. Our tri-fold phone is the first ever to come equipped with a 322-squarecentimeter UTG-glass display, offering every consumer a visual symphony in the palm of their hands.
• We also launched the first-of-its-kind Ultra Chroma Camera, which is equipped with 1,500,000 spectral channels to produce incredible true-to-life colors. This camera, along with the Ultra Lighting Camera, improved color accuracy by an astounding 120%.
• We continue to explore how to use AI to improve consumer experiences, including through applications like Pure Voice. This AIpowered feature can intelligently sense the caller, enhance their voice, and eliminate background and environmental noise for crystalclear calls even in the noisiest environments.
• We used AI Cloud Enhancement to improve photo quality and clarity, bringing users an ultimate 10x telephoto experience.
• We also unveiled the HUAWEI TruSense System, which allows our smart watches to accurately measure blood pressure through high-precision sensors.

Foundational Software 

• Our GaussDB became the industry's first database to be awarded the Evaluation Assurance Level 4 Augmented (EAL4+) certificate, both in China and globally, as part of the Common Criteria for Information Technology Security Evaluation (CC). It is also the only database to receive the highest level of security and reliability evaluation – Level II – from China.
• Thanks to breakthroughs in multi-dimensional awareness and scheduling, as well as subsecond in-place upgrades, the performance of openEuler improved by 15% in four major Kunpeng-powered scenarios, making it easier to upgrade cloud hosts and realize "zero migration".
• We also coordinated the optimization of different resources, including the Kernel, Compiler, OS Window System, and Application Framework, to better the performance of HarmonyOS 5 by 30% compared with the previous version.

Systems Engineering

• Detection and localization of silent data corruption in large-scale AI computing clusters have long plagued the industry. However, our breakthroughs in this area allow the detection of 90% more faults while still guaranteeing the performance of such systems.
• To address reliability challenges of HarmonyOS 5, we made breakthroughs in high-performance stack unwinding and concurrent fault localization technologies, achieving an accuracy rate of over 94% for the automated detection of reliability and performance faults.
• In software engineering, we built three lines of defense for HarmonyOS, which can identify more than 80% of security issues of software before they occur.