The peak power consumption does not exceed 110 watts, and the equivalent theoretical peak speed can reach 166.4 trillion fixed-point operations per second. On the afternoon of May 3, Cambricon Technology released China's first cloud smart chip in Shanghai - Cambricon MLU100 chip and board products, Cambricon 1M terminal smart processor IP product. Lenovo, Dawning and iFLYTEK, as Cambrian partners, have simultaneously released Cambrian chip-based application products.
As a pioneer in the field of international smart chips, on the eve of the explosion of artificial intelligence, Cambrian Technology once again brought new breakthroughs, and it also "leads human society from the information age to the intelligent age, and is a great supporter of the intelligent age." A step closer to the ideal of a chip company. It is worth mentioning that the MLU100 cloud smart chip is the result of nearly 2 years of research and development by 100 researchers in Shanghai Cambrian.
From the end to the cloud, Cambrian sets a new benchmark for AI chips
The focus of this conference was the Cambricon MLU100 cloud smart chip, which was officially unveiled for the first time.
Cambrian chip conference site
MLU100 adopts Cambrian's latest MLUv01 architecture and TSMC 16nm advanced process, it can work in balanced mode (1GHz main frequency) and high performance mode (1.3GHz main frequency), the equivalent theoretical peak speed in balanced mode is up to 1 second 128 trillion fixed-point operations, the equivalent theoretical peak speed in high-performance mode can reach 166.4 trillion fixed-point operations per second, but the typical board-level power consumption is only 80 watts, and the peak power consumption does not exceed 110 watts.
Like the Cambrian series of terminal processors, the MLU100 cloud chip still continues the excellent versatility of Cambrian products, and can support various deep learning and classic machine learning algorithms, fully satisfying vision, speech, natural language processing, and classical data. Cloud intelligent processing requirements in complex scenarios such as mining and other fields (such as large data volume, multi-tasking, multi-modality, low-latency, high-throughput).
The MLU100 board uses the PCIe interface, and its shape design is inspired by the ancient marine creature trilobites in the Cambrian geological era. Black and blue are the main colors, which are simple and full of technology.
It is worth mentioning that the MLU100 cloud smart chip is the result of nearly two years of research and development in Shanghai Cambrian, and it is the crystallization of Shanghai's wisdom. Cambrian landed in Lingang, Shanghai in April 2016, focusing on the research and development and sales of cloud-based smart chips. At present, Shanghai Cambrian has over 100 chip R&D personnel, with offices in Lingang and Zhangjiang, and an academician workstation based on Shanghai Brain Intelligence Project to carry out forward-looking academic research.
Device-Cloud Collaboration, Three Generations of IP Force "Intelligent Ecology"
The newly released Cambrian 1M is the company's third-generation IP product. The performance ratio of 8-bit computing under the TSMC 7nm process reaches 5Tops/watt (5 trillion operations per watt), providing three scales. The processor cores (2Tops/4Tops/8Tops) can meet the needs of different levels of intelligent processing in different application scenarios, and can further improve the performance through multi-core interconnection.
Dr. Chen Tianshi
In 2016, Cambrian launched the first-generation terminal intelligent processor IP product Cambrian 1A, the world's first commercial terminal intelligent processor IP product, which has been used in tens of millions of intelligent terminals (including Huawei Mate10, P20 and Honor 10 and other mobile phones).
The Cambrian 1M processor continues the excellent completeness of the first two generations of Cambrian IP products (Cambrian 1H/1A). A single processor core can support diverse deep learning models such as CNN, RNN, and SOM, and go further It supports classic machine learning algorithms such as SVM, k-NN, k-Means, and decision trees, supports local training, and provides a flexible and efficient computing platform for vision, speech, natural language processing and various classic machine learning tasks, and will be widely used. In different fields such as smart phones, smart speakers, smart cameras, and smart driving.
Both the Cambrian terminal and cloud products natively support the Cambrian NeuWare software tool chain, which can facilitate the development, migration and tuning of smart applications.
Conference site. (Photo by Xie Zhenlin)
Dr. Chen Tianshi said that the original intention of Cambrian was to make intelligent processors available to the whole world. Cambrian will adhere to the spirit of openness and collaboration in the academic community, and share the latest technological achievements of Cambrian with its counterparts around the world in the form of processor IP authorization, enabling global customers to rapidly design and produce chip products with artificial intelligence processing capabilities.
Cambrian implements the concept of "device-cloud collaboration" in technology. The MLU100 cloud chip released this time can not only complete various complex cloud intelligent tasks independently, but also work with Cambrian 1A/1H/1M series terminal processors. Perfect adaptation, allowing the terminal and the cloud to work together to complete complex intelligent processing tasks on the basis of a unified intelligent ecosystem.
Cambrian ecological outbreak, industry partners show chip-related applications
After Dr. Chen Tianshi's keynote speech, some Cambrian industry partners publicly demonstrated the application solutions based on Cambrian chips.
Lenovo Group launched the ThinkSystem SR650 based on the Cambricon MLU100 intelligent processing card at today's press conference, breaking the world record of 37 server benchmark tests. This product is a 2U2-channel rack-mountable specification and supports 2 MLU100 intelligent processing cards. The launch of this cloud intelligent server will fully support Lenovo customers' business needs in machine learning/VDI/virtualization/cloud/database/analytics/SAP.
Chen Tianshi and his mentor Chen Guoliang jointly released new products
Sugon also launched the server product series "PHANERON" based on Cambricon MLU100 intelligent processing card at today's press conference.
The PHANERON series was launched in November 2017 and is based on a prototype test chip previously announced by Cambrian. The new "PHANERON" released this time has been upgraded synchronously with the release of the MLU100 smart accelerator card, and has stronger performance. It can support 2-10 Cambrian MLU processing cards and flexibly respond to different smart application loads.
Taking the upgraded version of PHANERON-10 as an example, a single server can integrate 10 Cambrian artificial intelligence processing units, which can provide 832T half-precision floating-point computing power for artificial intelligence training applications, and 1.66P integer computing power for inference applications. The energy efficiency in typical scenarios is improved by more than 30 times. In addition, Sugon also disclosed the artificial intelligence management platform SothisAI, which can achieve seamless connection and deep integration with Cambrian chips and its development environment.
iFLYTEK, the leader in intelligent voice applications, also disclosed its in-depth research and development project with Cambrian chips at today's press conference. According to the introduction, it takes 10,000 hours to complete one hour of voice data for intelligent application processing on a traditional processor. Therefore, iFLYTEK has been tracking the cutting-edge progress of AI-specific chips. As early as 2014, iFLYTEK began to communicate with the early Cambrian research team on the implementation of related speech algorithms on the processor, and invested in the Cambrian in the Angel Wheel in 2016, and began to use the Cambrian chip in 2017. Test the application on.
The test results disclosed by iFLYTEK at the press conference show that the Cambrian’s intelligent processor has delivered an excellent answer in voice intelligent processing, and its energy consumption efficiency is more than 5 times higher than that of the competitor’s cloud GPU solution. The powerful processing capability of the Cambrian intelligent processor enables the local end of the mobile phone to process more complex machine learning algorithms, making the local speech recognition accuracy 9.8% higher than that of traditional processors, significantly improving the user experience.
Source: "Wen Wei Po"
Recommended reading
Blockchain report | Brain-computer interface report
Silicon Valley Artificial Intelligence | Stanford Chancellor
Wei Zhe | Yao Jinbo | Hu Haiquan
Vertical Planting | Unmanned Vehicles