# Research & Innovation

Scientific exploration and technological innovation drive civilization and society forward. Huawei understands the importance of research and innovation and how openness is critical for both. We are ready and willing to work with academia and industry to explore the frontiers of science and technology, push innovation forward, create value for industry and society as a whole, and build a better intelligent world.

## Basic Research: Overcoming Theoretical and Industry Bottlenecks with Patient Investment into Communications and Computing

### Mathematics is the queen of the sciences and key to the new world of communications, computing, and AI. For this reason, Huawei consistently conducts research into mathematical theory and its applications.

In the field of random matrix theory and algorithmics, we created a new method of estimating the minimum singular value of an n × n random matrix of a sparse linear system by running an algorithm in $$\widetilde{O}\left(n^{\frac{3 \omega-4}{\omega-1}}\right)$$ time where ω is the matrix multiplication exponent. The paper that elaborates on this method was accepted in 2022 by the Annual ACM Symposium on Theory of Computing (STOC), the flagship conference of the Special Interest Group on Algorithms and Computation Theory

In terms of training deep learning models, we proposed the Boolean logic backpropagation method. This breakthrough in mathematical principles helps solve problems related to hyperdimensional binary variable optimization and makes it possible to both conduct deep learning algorithms in the Boolean field and realize the performance of a full-precision model through simulated verification.

In the field of topos theory and semantics, we put forward a brand-new mathematical theory for semantic learning, creating a public framework for algebraic topology and logical propositions. We also proposed the world’s first semantic information spaces based on groups, partially ordered sets, and groupoids, which explains well the experimental results of Roger N. Shepard, Carl I. Hovland, and Herbert M. Jenkins. The experiment shows that the new neural network model is not just a “statistical parrot” as it can reach a more animal-/human-like intelligence.

### We have also dived deep into communications and computing theories, constantly exploring, challenging, and approaching the theoretical limit.

In the field of information theory, we put forward a coding method based on multilateral coupling graphs. This method allows us to increase performance by about 0.5 dB in optical communications scenarios at a Tbit/s-level speed over thousands of kilometers, bringing us closer to the Shannon limit.

We have consistently pushed the limits of entropy coding with our unstructured data coding algorithm HZU, through which we discovered the fast nonlinear transformation and lightweight context prediction method. This algorithm can generate results superior to those of the Lempel-Ziv (LZ) compression paradigm and boost the compression ratio by about 30%.

In the field of network optimization, we created a topology-aware network pricing policy (NPP) model and a self-adaptive sparse optimization method to solve problems surrounding flows in networks that possess hundreds of thousands of nodes and tens of millions of links, improving problem-solving speeds by orders of magnitude. We also proposed the CLIModel based on the biological assimilation mechanism, improving network configuration efficiency by over five times. The paper on this model has received a Best Paper Award from ACM SIGCOMM 2022.

In the field of computing and AI, we introduced an allocation and scheduling method capable of minimizing communications costs and maximizing task-level parallelism in a weighted manner. Based on this method, we built Fold3D, a three-dimensional parallel training framework that delivers performance 25% higher than mainstream industry systems.

In the field of machine learning, we were the first to prove that analytical solutions existed for the optimal Kullback-Leibler (KL) divergence and optimal reverse variance of the diffusion probabilistic model (DPM). Such a breakthrough has the potential to improve sampling efficiency by 20- to 80-fold, and therefore the related research paper has been presented the ICLR 2022 Outstanding Paper Award.

## Technological Innovation: Driving Industry Development to Usher in a New World of 10 Gigabit Experience, 100 Billion Connections, and Native Intelligence

### Wireless Networks

We have been working with the global industry to explore and define 6G, and have also proposed six technology pillars for 6G.

We place equal importance on theoretical innovation and prototype verification, and have unveiled the first 220 GHz terahertz communications prototype system with a peak rate of 240 Gbit/s.

In terms of short-distance communications, we used mmWave to realize superb user experience featuring throughput of over 10 Gbit/s and a sub-millisecond latency.

We created an innovative hybrid antenna array architecture that combines both reflective and reconfigurable surfaces, achieving a coverage gain of nearly 41 dB in the 10 GHz band in non-line-of-sight (NLOS) scenarios.

We completed the world’s first field trial verification of mobile access to low-earth orbit satellites based on 5G NR protocols, and also became the first to simultaneously realize high-speed communications and millimeter-level high-precision imaging in the terahertz frequency band while maintaining the same hardware and the same waveform.

### Optical Networks

We continued our work to increase the single-wavelength rate for long-haul transmission from 400 Gbit/s to 800 Gbit/s by diving into key technologies like low-complexity nonlinear waveform design for high-speed signals and 200-GBaud ultra-high-speed optoelectronic modulators.

We also made breakthroughs in ultra-wide-spectrum S + C + L band optical amplifiers, next-generation higher-order modulation algorithms, and improved the performance equalization of wide-spectrum optical systems, expanding optical spectrum by 50%.

During a joint on-site test with Orange, we set a new record of single-fiber 157 Tbit/s transmission over a distance of 120 kilometers.

The joint test team that set the record, with members from both Orange and Huawei Optical Communication Technology Lab in Paris.

### Data Communication

Our latest bus-level data center network (DCN) uses a novel architecture to enable the convergence of the Ethernet, Fibre Channel (FC), and InfiniBand (IB), achieving a breakthrough in high-performance computing network and storage network performance.

We have continued to build next-generation Ethernet technology systems. By supplementing digital circuits with analog ones, we’ve reduced energy consumption by around 50%. By separating error detection from error correction, we’ve reduced interface latency by nearly 70%. By adopting flexible slicing, we’ve created private lines at n x 10 Mbit/s.

### Intelligent O&M

We have constructed an O&M pretrained foundation model (PFM) that unifies the representation of network element data and prevented siloed modeling for O&M. Compared with the previous models, each of which supported only one task, this new model can support multiple downstream tasks, such as fault data analytics, root cause analysis (RCA), responses to customer complaints, intelligent signaling analysis, and assistance for changes, while also greatly improving their efficiency.

By employing a technology of optimizing ultra-large-scale wireless parameters in parallel, we have been able to increase the cell-edge rate of 5G downlink users by more than 10%.

### Cloud and Computing

As part of our work to address the performance and latency challenges presented by cross-chip and cross-server communications in heterogeneous computing, we have redefined the bus architecture for computing cluster communications and unified the interconnection standards between chips and nodes. This means the live migration of virtual machines can be performed in mere milliseconds, substantially boosting network efficiency.

We launched the industry’s first heterogeneous computing cluster software platform, improving computing resource utilization by over 35% and accelerating performance by more than 100% in scenarios like deep big data analytics and retrieval.

We launched our new mobile photography brand – XMAGE – to bring new portrait and night mode features to consumers through innovative mobile photography technologies.

We’ve designed a unique double-rotating falcon wing hinge which, when combined with ultra- high-strength steel and flexible materials, makes foldable phones more durable, significantly lighter, and enables them to retain their flat screens, regardless of how many times you fold them.

By continuing our research into materials and production methods, we’ve made breakthroughs that have revolutionized phones. The sturdy Kunlun Glass has made our phones 10 times more drop-resistant, and one of our latest phone models, which was reinforced with Kunlun Glass, became the first to ever earn the five-star glass drop- resistance certification from Switzerland’s SGS.

### AI Algorithms

We launched ZooD, a model library technology that supports search functionality. Powered by Ascend, MindSpore, and Huawei Cloud, this breakthrough improves model performance by over 30%. Our OptVerse AI Solver also made its debut in 2022, making it possible to solve optimization problems involving tens of millions of variables and constraints in multiple scenarios.

For the first time, we realized compression via quantization of generative models, increasing their compression ratios by more than 10 times without affecting performance. The upgrade of the parallel distillation technology achieves a 100-fold increase in quantization speed, better supporting the deployment of Huawei Cloud’s full-precision models that have hundreds of millions, tens of billions, or even hundreds of billions of parameters.

### Foundational Software

We are continuing to patiently invest in foundational technologies that will deepen our roots in the software industry.

By innovating on microkernel, storage, and scheduling architectures for operating systems, we have significantly increased the utilization of hardware resources.

The groundbreaking accomplishments we have made in fusion storage engines, fully-encrypted databases, and AI-based optimizers through the introduction of GaussDB have sharpened our competitive edge in performance, security, and availability.

We have remained committed to an open source strategy and keeping an open mind while building software ecosystems. Through our cooperation with players across the industry, openEuler, openGauss, and other software ecosystems are now on track to undergo rapid development.

• The openEuler, openGauss, and MindSpore ecosystems have rallied more than 2,000 partners and over three million developers, with more than three million openEuler installations to date.
• HarmonyOS had been deployed on 330 million Huawei devices. This growth has been supported by our collaboration with over two million developers and more than 2,300 HarmonyOS Connect partners within the HarmonyOS ecosystem.

### Security and Trustworthiness

We remain dedicated to the research of technologies related to security and trustworthiness and work hard to improve engineering technologies and capabilities, such as software engineering and systems engineering in the R&D domain, so as to build secure, resilient, reliable, and quality products for customers.

We have built an end-to-end IPD-SFP cyber security engineering system which covers R&D, procurement, and delivery and services. By institutionalizing best practices and incorporating them into business processes and IT operating platforms, we are developing advanced cyber security capabilities.

We have created a digital system design methodology that is centered on full digital design assets and allows massive numbers of documents to be passed on or transferred. With this method, empirical assets in large-scale, complicated systems can now be reused more efficiently, increasing the agility of design and verification.

Huawei is an active member of and major contributor to open source communities like Rust and standards organizations such as ISO, ITU-T, 3GPP, and IETF. Together, we are laying the cornerstone of trust in the digital world.

## Exchange, Enlighten, Explore: Going Beyond Physical Boundaries to Ignite the Spirit of Inquiry and Academic Exchange

Scientific exploration is an undertaking that hinges on communication, sharing, cooperation, and patient investment. Scientific research is most powerful when driven by close engagement and collaboration between scientists and experts. Huawei remains open to collaboration and is always looking to join hands with outstanding researchers and partners from around the world to address the biggest challenges of our times.

In 2022, Huawei officially launched the Chaspark Technology Website, which has attracted more than 120,000 authenticated users from across multiple academic fields. We intend to build Chaspark into a world-class, open platform for academic exchanges. To date, the platform has presented over 7,000 hot topic posts and research papers, more than 2,200 academic activities, and 285 technical challenges, allowing experts from outside Huawei to propose their own solutions.

June 2022 saw the first issue of the quarterly journal Communications of Huawei Research , which documents and shares the latest findings of researchers, both inside and outside of Huawei, from across the various research domains in which the company is engaged. By December 31, 2022, we had published three issues, including 39 academic papers and research findings from over 170 Huawei scientists and external researchers.