“According to Bloomberg data, the global AI market size will be about $86.9 billion in 2022, and this figure is expected to climb to $1.3 trillion by 2030.” These are the numbers shared by Keysight at the Keysight World Tech Day 2025 media conference in Shanghai on June 26.
As a hard-core technology event with weather vane significance, this year’s Keysight World Tech Day has built a high-end communication platform for all sectors of the industry to discuss the latest industry trends and cutting-edge technologies through a summit forum and three parallel sub-forums: “AI-driven ultra-high-speed transmission testing”, “B5G/6G wireless cutting-edge technology and testing”, and “comprehensive chip testing from power to AI”.
When asked why AI and B5G/6G were chosen as the focus of this conference, Marie Hattar, senior vice president and chief marketing officer of Keysight, said, “The three areas of 6G, AI and IC have great growth potential. “These technologies are in a period of upgrading and require advanced test and measurement solutions. Keysight, as an established manufacturer deeply involved in the field of test and measurement, hopes to cooperate with the industry and actively participate in the development of solutions and the formulation of industry standards.
AI technology landing: data centers “sick to work”andKAI Solution
AI is not a new concept. The AI market we are discussing today often focuses on generative AI. Early AI mainly focused on pattern recognition, showing unidirectionality, and required relatively small computing power, parameters, and data scale. Today’s generative AI is no longer just about simple identification, but requires larger models to ensure that the final generated content is more accurate. Therefore, the scale of AI model parameters has now risen to trillions. At the same time, computing power clusters have evolved to 10,000 cards or even 100,000 cards.
Such computing power is very expensive. If you want to build a true data center, the investment is often as high as tens of billions of yuan, including GPUs, CPUs, TPUs, storage devices, various active and passive cables, etc. Among these investments, the cost of GPUs is undoubtedly one of the highest. However, our findings show that GPUs aren’t “working” all the time.
When training the model, the data showed that the GPU had a waiting time of more than 50%. Further analysis found that GPU accounts for 20% of the computation time, 16% of the overlap time, and 2% of the memory time during model training. The remaining 62% of the time the GPU is in a waiting state.
Why are GPUs always waiting? There is a simple logic behind this: if the GPU does not receive data, it will naturally not be able to perform calculations. So, why didn’t the data reach the GPU in time? Li Jian, manager of high-speed digital marketing at Keysight Greater China, explained: “It’s like encountering a traffic jam on the way to work, then we will be late for work, and we can’t start work on time. There are also various obstacles when data exchange and transmission. ”
In addition to the low utilization rate of core assets in data centers, Li Jian also pointed out another situation worth paying attention to – frequent failures in data centers.
The data shows that the probability of successfully completing the training task in one go is only about 57%. This is because during training, there are various hardware or software failures, especially hardware failures.
Compared with small models and small and medium-sized models in the past, the current “large model” requires the support of large computing power, big data, and large bandwidth. Nowadays, the rate is increasing very quickly, and many devices are approaching their physical limits. In a high-temperature, high-speed, and 24-hour working environment, many devices are actually overwhelmed or even collapsed. If any of these devices fail or have problems, the entire system may not function properly and must be shut down, restarted, or restored. If you don’t take some specific measures, such as breakpoint protection and breakpoint restart, your data may be scrapped even if you spend a lot of time training before.
At the heart of this problem is that the foundational hardware and software, which are the cornerstones of the data center, are not reliable. Many computing power centers are actually “online with illness”, and there are many problems. Can the huge amount of money invested in the construction of data centers and computing power centers meet expectations? Is there still room for optimization? Li Jian’s answer is “there is a lot of space”, and Yete Technology also pays close attention to industrial development, hoping to jointly improve the indicators of grassroots hardware and software in the data center, detect various problems that may be hidden in the network in a timely manner, dredge network traffic, and make the network operation more effective. This is why Keysight has gradually sorted out its product line in the past two or three years, redefining the product matrix of artificial intelligence for the artificial intelligence market, and unified it as Keysight AI (KAI).
Keysight divides the product matrix into four sections:
· KAI high-speed computing: The bottom layer of the data center is the computing power board, which contains a large number of high-speed devices, including computing power devices, memory devices, and various interfaces and traces. This part provides all the testing methods associated with the computing power board to help create more stable and reliable computing power.
· KAI interconnection: Computing power boards need to be connected using active or passive cables, optical modules, optical connections, etc. to form nodes or even supernodes. The number of these devices operating at high speeds is large, so quality must be guaranteed. Keysight provides network analyzers, sampling oscilloscopes, high-speed bit error meters, and more.
· KAI network: When supernodes are formed into a larger computing power network through network cards, switches, etc., a lot of testing is required on the network side. Keysight leverages the acquired Ixia technology to form the main solution in this segment.
· KAI Energy Efficiency: “The end of the data center is the power station”. A computing power center with 100,000 boards consumes electricity comparable to or even exceeds the power consumption of a city with a population of one million. A large number of high-density devices are used in data centers, ranging from AC to DC, DC to DC, etc. The power conversion efficiency of power devices is directly related to the hardware, investment, and operating costs of computing power centers. Keysight offers a wide range of KAI energy efficiency solutions.
Li Jian introduced: “We hope to support the whole life cycle of user products, starting from the pre-research stage, through product research and development, production, and then to installation and maintenance, to meet the testing needs of users at all levels from physical layer verification, to network layer, and then to application layer. ”
Keysight has more than 55 solutions in the field of AI, covering more than 110 different infrastructure components. At the Optical Network and Communications Symposium and Expo (OFC) in March and April this year, Keysight released three new products corresponding to the application scenarios of the physical layer, network layer and application layer.
· DCA-M Sampling Oscilloscope: 224G single-channel and dual-channel sampling oscilloscopes meet the needs of leading manufacturers in the order of 224G, targeting optical modules, optical chips and component manufacturers. The core device of the sampling oscilloscope is independently developed by Keysight High Frequency Chip Technology Center (HFTC), using indium phosphide technology, which performs well in core indicators such as noise and jitter.
· Interconnect and network performance test equipment: Perform the verification of the transport layer and network layer after the optical devices, optical modules and network cards, switches, and various other modules are connected. The product has a speed of up to 1.6T and has a layer 2 forward error correction coding function, which can perform dynamic testing and extreme performance testing.
· KAI Data Center Builder: Compared with traditional data centers, today’s computing power centers generate a large amount of “east-west traffic”. If the traffic throttling mechanism is not perfect, the GPU will spend a lot of time waiting for data. The key is to simulate the various traffic generated by different large model operating environments. This product simulates various traffic conditions that may occur at various nodes in a real network, especially pulsed traffic. Keysight is the first company in the industry to do this.
6GEvolution: inAILooking for the “profit code” of the communications industry in the wave
“6G is not about building a new network that completely replaces 5G.” According to Bai Ying, manager of the wireless marketing department of Keysight Greater China, the global 5G has not yet completed full coverage, and 6G needs to “innovate in inheritance”.
People have high hopes for “killer” applications. In the 2G era, people send text messages, in the 3G era, people begin to use the network, 4G mobile Internet explodes, and for 5G, people will put their hopes on the 5G private network. However, due to industrial barriers and high costs, its development is not satisfactory.
When talking about the combination of 6G and AI, Bai Ying believes that “AI will definitely be applied to 6G networks, but the ultimate question we need to ask is, what can the industry get? What benefits can it bring to end consumers and enterprise-level consumers? What improvement has been made to people’s work? Or can it save more costs for enterprises? Will the experience be better for the end user? I guess it’s a very real question. ”
He explains the value of technology beyond generational labels: “What makes consumers really accept new things is the key. We can call it 6G, the name is not important, but it must represent something that 5G cannot do. ”
Keysight has always followed industry standards and hopes to be one of the contributors to the development of standards. Keysight uses software technology to automate intelligent tests, which is also part of AI in a sense. For example, in Madagascar, Keysight and operators built a virtual digital twin based on geographic information knowledge, so that operators can choose sites reasonably according to changes in the flow of people without the need for on-site surveys when setting up stations.
In the new product and solution showcase, Keysight World Tech Day focused on a series of innovative technologies, cutting-edge applications and market hotspots to customers and partners in the “B5G/6G Paves the Way for Wireless Technology Leap”, including:
FR3 Band Test Solution: Based on Keysight’s VXG signal generator and UXA signal analyzer, it provides up to 4GHz broadband signals for full coverage of the FR3 band. It has leading EVM, ACP and flatness performance indicators, and supports the generation and analysis of user-defined signal formats.
Low-phase noise source for 6G communication applications: AnaPico’s economical and compact test solution features ultra-low phase noise, ultra-fast switching speed, wide frequency range, channel-to-channel coherence, and supports AM, FM, PM, and pulse/IQ signal modulation.
Wi-Fi 7 Signaling Test Solution: The E7515W UXM wireless test platform supports frequencies up to 7.125GHz, capable of performing protocol, RF, and functional tests on Wi-Fi STA devices and AP devices, and can provide test support for the entire process of Wi-Fi products from early design to acceptance and deployment.
