October 16, 2024 - At the SEMiBAY2024 "HBM and Memory Technology and Application Forum", Xiong Dapeng, founder, chairman and CEO of Yizhu Technology, delivered a speech entitled "Beyond the Limits: Technical Challenges and Solutions Facing High-Computing Power Chips".
Dr. Xiong Dapeng proposed that driven by AI big model technology, computing power is reaching a turning point in demand, and hardware architecture will become one of the key paths to meet computing power demand. Future computing power growth will be centered on storage units.
Opportunities and Challenges in the Era of Big Models
In the era of AI big models, with the continuous improvement of data, computing power, and parameter volume, model capabilities have been significantly enhanced. Dr. Xiong Dapeng pointed out that big models have gradually evolved from quantitative changes to qualitative changes. When the model size is large enough, emergent capabilities similar to human "enlightenment" will appear, and the reasoning ability of big models will be significantly improved. This change indicates that the last mile of AI application is about to be opened up, and the implementation of business will drive the demand for AI computing power to a turning point.
Omdia's latest report, "Cloud Computing and Data Center Artificial Intelligence Processor Forecast", shows that the market size of GPUs and other acceleration chips used for cloud computing and data center artificial intelligence has grown from less than $10 billion in 2022 to $78 billion in 2024, and is expected to reach $151 billion by 2029. However, the market may see a clear inflection point in 2026, and the growth momentum will shift from technology adoption to changes in demand for artificial intelligence applications.
In addition, IDC predicts that future AI servers will focus on improving computing power and processing efficiency (energy efficiency ratio) to adapt to more complex and larger-scale AI applications. It is expected that by 2027, the proportion of AI computing power used for reasoning will reach 72.6%, and in the future it is expected to reach 95% for reasoning and 5% for training.
Application implementation requires hardware architecture breakthroughs
However, the speed of improving the performance of existing chip hardware can no longer meet the rapidly growing computing power requirements of algorithm models. Moore's Law, the golden rule that once guided the development of the semiconductor industry, is now facing unprecedented challenges. A report from the Economic Research Institute of Guosen Securities pointed out that the parameter scale of large models increases 35 times every 18 months, while chips under Moore's Law only increase by 2 times. Therefore, exploring and developing new hardware architectures has become one of the key paths to breakthroughs in computing power.
Dr. Xiong Dapeng emphasized that under the existing hardware architecture, AI chips are currently facing the "three wall" problem: storage wall, energy consumption wall and compilation wall. The storage wall refers to the problem that the data access speed of the memory cannot keep up with the data processing speed of the computing unit, resulting in a performance bottleneck.
At the same time, the existence of the storage wall brings about the problems of energy consumption wall and compilation wall. The energy consumption wall refers to the fact that as chip performance improves, energy consumption and heat dissipation issues become the main factors limiting further performance improvement. The compilation wall refers to the fact that as the complexity of AI models increases, the amount of data and computing tasks that the compiler needs to process also increase dramatically, which makes static compilation optimization very difficult, and manual optimization consumes a lot of time and cost.
Storage and computing integration opens up the second growth curve of computing power
Faced with this challenge, Yizhu Technology chose to innovate and used a new chip design concept, the "storage-computing integrated super-heterogeneous" architecture, which greatly reduced the delay in data transfer and improved the overall computing efficiency and energy efficiency.
Dr. Xiong Dapeng pointed out that if we want to break the "three walls" of AI chips, we need to start from the first principle of computing power (Amdahl's law) and significantly reduce the amount of data moved so that the F value is close to 0, so as to ensure the linear growth of effective computing power density. Currently, there are two main solutions in the industry: one is in-memory computing, and the other is near-memory computing.
In-memory computing integrates storage and computing functions to reduce data transfer latency and improve performance and energy efficiency. In an ideal state, F=0, which enables seamless integration of storage and computing. Near-memory computing integrates storage units and computing units through advanced packaging to increase memory access bandwidth, reduce data transfer latency, and improve overall computing efficiency.
Dr. Xiong Dapeng emphasized that through technologies such as storage-computing integrated architecture, we can break through the bottleneck of traditional computing models, achieve higher effective computing power, and break the ceiling of effective computing power. In the future, the era centered on computing power units will come to an end, and the second growth curve of computing power will be centered on storage units.
Conclusion
Dr. Xiong Dapeng said that since its establishment, Yizhu Technology has always been committed to providing a new path for the development of AI high-computing chips that are more cost-effective, more energy-efficient, and have greater computing power development space through storage and computing integration. In March 2023, in the face of AI computing challenges brought by large models such as ChatGPT, Yizhu Technology first proposed "storage and computing integrated super heterogeneity", providing a new idea for the development of AI high-computing chips in the era of large models.
In the future, with the continuous advancement of AI technology, the demand for computing power is also growing. Yizhu Technology will provide a new direction for the development of AI chips through an innovative storage-computing integrated architecture. In the era of large models, Yizhu Technology's technology and products will provide strong support for the development of AI technology and drive the entire industry forward. With the continuous maturity of Yizhu Technology's technology and the continuous expansion of its applications, we have reason to expect that AI chip technology will usher in a new stage of development and make greater contributions to scientific and technological progress!
Previous article:Akamai adds behavioral DDoS protection engine to App & API Protector
Next article:Edge AI: Revolutionizing Real-Time Data Processing and Automation
- Popular Resources
- Popular amplifiers
- Keysight Technologies FieldFox handheld analyzer with VDI spread spectrum module to achieve millimeter wave analysis function
- Qualcomm launches its first RISC-V architecture programmable connectivity module QCC74xM, supporting Wi-Fi 6 and other protocols
- Microchip Launches Broadest Portfolio of IGBT 7 Power Devices Designed for Sustainable Development, E-Mobility and Data Center Applications
- Infineon Technologies Launches New High-Performance Microcontroller AURIX™ TC4Dx
- Rambus Announces Industry’s First HBM4 Controller IP to Accelerate Next-Generation AI Workloads
- NXP FRDM platform promotes wireless connectivity
- WPG Group launches Wi-Fi 7 home gateway solution based on Qualcomm products
- Exclusive interview with Silicon Labs: In-depth discussion on the future development trend of Bluetooth 6.0
- Works With Online Developer Conference is about to start, experience the essence of global activities online
- LED chemical incompatibility test to see which chemicals LEDs can be used with
- Application of ARM9 hardware coprocessor on WinCE embedded motherboard
- What are the key points for selecting rotor flowmeter?
- LM317 high power charger circuit
- A brief analysis of Embest's application and development of embedded medical devices
- Single-phase RC protection circuit
- stm32 PVD programmable voltage monitor
- Introduction and measurement of edge trigger and level trigger of 51 single chip microcomputer
- Improved design of Linux system software shell protection technology
- What to do if the ABB robot protection device stops
- CGD and Qorvo to jointly revolutionize motor control solutions
- CGD and Qorvo to jointly revolutionize motor control solutions
- Keysight Technologies FieldFox handheld analyzer with VDI spread spectrum module to achieve millimeter wave analysis function
- Infineon's PASCO2V15 XENSIV PAS CO2 5V Sensor Now Available at Mouser for Accurate CO2 Level Measurement
- Advanced gameplay, Harting takes your PCB board connection to a new level!
- Advanced gameplay, Harting takes your PCB board connection to a new level!
- A new chapter in Great Wall Motors R&D: solid-state battery technology leads the future
- Naxin Micro provides full-scenario GaN driver IC solutions
- Interpreting Huawei’s new solid-state battery patent, will it challenge CATL in 2030?
- Are pure electric/plug-in hybrid vehicles going crazy? A Chinese company has launched the world's first -40℃ dischargeable hybrid battery that is not afraid of cold
- Millimeter wave radar + smart elderly care =?
- msp430 program library serial series asynchronous serial port USART
- A look inside a "high-end" gas detector
- Verify the PEDOMER (pedometer) function through the APP
- Online text and code difference comparison tool
- Please, please take a look at this. A beginner can't do it at all.
- Oscilloscope measuring the waveform of BYD Qin's three-phase drive motor
- C5000 compiles SUBC instruction to implement division
- Are you interested in CNC power supply?
- 122X32LCD screen printing CBG122032AL00 driver SED1520 screen data