Deeply engaged in the field of AI chips, Kunlun Core Technology helps build an AI chip ecosystem integrating software and hardware

The best AI chip of 2022

Best Entrepreneurial Team of 2022

The Most Innovative IC Design Company of 2022

Company Introduction

https://www.kunlunxin.com.cn

product description

① Adopting the new generation of self-developed architecture Kunlun core XPU-R, the versatility and performance are significantly improved.

② New generation self-developed architecture: Using Kunlun core XPU-R architecture, the versatility and performance are significantly improved.

③ Powerful computing power: 256 TOPS@INT8, 128 TFLOPS@FP16

④ Technological leadership: 7nm advanced process, GDDR6 high-speed video memory.

⑤ Complete functions: supports hardware virtualization, inter-chip interconnection and video encoding and decoding; 7nm advanced technology 32GB high-speed memory 512GB/s memory bandwidth.

Kunlun Core 2nd Generation Chip is based on the software-defined AI chip architecture Kunlun Core XPU-R, which has the following core advantages:

General computing capabilities have been significantly enhanced, which can flexibly support the evolution of AI algorithms and improve the effectiveness of resource investment.

The self-developed XPU-R architecture increases the general computing core computing power by 2-3 times, greatly enhancing the general computing capabilities of the product. For typical AI loads, the measured throughput performance of the R200 AI accelerator card is 1.5 times that of the industry's mainstream 150W GPU.

Hardware virtualization improves the utilization of AI computing resources

Supports hardware virtualization function, and its computing unit and storage unit can be physically isolated and used by multiple users. While ensuring the Quality of Service (QoS), the resource utilization of the AI ​​accelerator card is significantly improved.

High-performance distributed AI system, accelerating high-speed data exchange in AI data parallelism and model parallelism

Kunlun Core R480-X8 AI accelerator group adopts OAM module to provide AI computing power of up to 1PFLOPS FP16 for single-node AI servers. At the same time, through multi-chip interconnection, the product can provide 200GB/s aggregate bandwidth, which can effectively support the high-speed exchange requirements of data in training strategies such as model parallelism and data parallelism.

Kunlun Core Technology continues to optimize underlying core technologies such as chip architecture and instruction sets to adapt to artificial intelligence applications and various algorithms, and continuously improve product performance, energy efficiency and ease of use. At present, Kunlun Core Technology's products are benchmarked with international mainstream solutions in terms of parameters, which can provide better performance, power consumption ratio and cost performance.

Taking the R200 AI accelerator card as an example, after actual testing of business-scale deployment, the performance improvement for typical AI loads is about 1.5 times. Taking the throughput rate in the inference scenario as an example, the acceleration effect is as follows:

-GEMM general matrix multiplication performance is accelerated by 1.7 times;

-BERT: Bert, a typical algorithm for natural language processing, has a performance acceleration of 1.4 times, and has excellent acceleration performance for other Transformer-like algorithms;

-YOLOv3 visual target detection algorithm YOLO performance acceleration is 1.3 times;

Kunlun Core 2nd Generation Chips are mainly aimed at the high-performance data center inference market, including cloud and edge data centers. They flexibly support deep learning and machine learning algorithms such as vision, voice, natural language processing and search, and flexibly support user-defined operator development. Kunlun Core 2nd Generation Chips support mainstream Internet applications, pan-vision, finance, industrial Internet, government affairs and other industry applications.

Kunlun Core's first-generation products have been deployed in more than 20,000 pieces in Baidu search engine, Xiaodu and advertising businesses, and have more than 50 external customers. It is the only AI chip in China that supports large-scale core algorithms of the Internet. After passing through the Internet data center The most stringent business launch test has fully verified the product's availability, reliability, stability, and robustness. It also proves the technical strength of the Kunlun Core team in chip architecture, software stack, and system engineering.

Application case one: Baidu search engine

The business forms of data centers are diverse and have different requirements for AI algorithms. For example, Baidu’s search engine business is mainly based on natural language understanding NLP, supplemented by vision and voice. This requires AI accelerator cards to support multiple types of AI algorithms truly provide universal support for business algorithms. Kunlun Core's products have been deployed at the Baidu data center at the 10,000-ka level and are currently running stably.

Data center search has strong requirements for business real-time and high concurrency. Compared with mainstream GPU inference cards, Kunlun Core's accelerator cards can provide higher performance and lower costs, and the overall TCO is reduced by more than 1/3. Achieved commercial cost reduction and efficiency increase.

Application Case 2: Industrial Machine Vision

The industrial quality inspection solution of Kunlun Core 2nd Generation Chip has been applied on a large scale in a domestic intelligent manufacturing enterprise, achieving quality inspection of hundreds of millions of 3C parts of domestic and internationally renowned brands, completing the maximum replacement of manual quality inspection, greatly saving labor costs, and the overall solution can recover costs in about 14 months. At the same time, Kunlun Core products support the overall solution of "5G+AI+Industrial Internet", which can greatly improve the intelligence level of traditional enterprises, help enterprises reduce losses, increase the yield rate by about 10%, and increase corporate profits.

Application case three: localized financial business

Currently, there is a large amount of document image data in banking and other financial services that needs to be extracted manually, which creates huge efficiency and cost bottlenecks. At the same time, financial IT localization level indicators require the use of nationalized solutions when meeting business needs. . Kunlun Core 2nd generation chip supports a large number of mature and reliable commercial OCR models and algorithms. The hardware is equipped with an all-in-one machine based on domestic CPUs, which can quickly connect to customer business systems to achieve accurate structured data extraction such as identity documents, improving business execution efficiency. This solution was successfully implemented in a domestic commercial bank in 2021, becoming the first domestically produced AI solution to implement an AI capability engine.

Vote to win prizes

Click to read the original text and vote for the company↓↓↓