Suyuan Technology released its first AI training product, "Cloud Sui T10" | Zhongji Family
On December 11, 2019,燧原科技 (Suyuan Technology) released its first AI training product, "Cloud燧T10". This is an AI training accelerator card for cloud data centers, built on the high-performance general-purpose AI training chip "邃思". It can be widely used in cloud training scenarios for the internet, finance, and government sectors. At the same time, Suyuan Technology also released "驭算", a new computing and programming platform that supports mainstream industry frameworks.
Zhang Yalin, COO of Suyuan Technology, introduced: “The newly released ‘Cloud燧T10’ boasts industry-leading performance indicators. We independently developed a fully general-purpose programmable neural network processor and overcame the challenges of 2.5D packaging, entering the product trial and commercial use phase within 18 months. Notably, to address the challenges of large-scale training clusters in cloud data centers, we creatively proposed a 200GB bi-directional ESL interconnect technology, significantly reducing system complexity and cost. The simultaneously released 驭算 platform provides AI developers with a convenient and user-friendly programming environment. In addition, we have also launched the ‘Cloud燧T11’ based on the OCP Accelerator Module (OAM).”
“Cloud燧T10” is a dual-slot card with excellent computing power, versatility, and value features:
Exceptional Computing Power: 20 TFLOPS in single-precision (FP32); 80 TFLOPS in half-precision and mixed-precision (BF16/FP16), with a maximum power consumption of only 225W, leading in energy efficiency;
Model Versatility: Supports common AI training models such as CNN, RNN, LSTM, and BERT, applicable to image, stream data, and speech training scenarios;
High-Speed Interconnect: Utilizes Suyuan ESL interconnect technology to meet the deployment needs of E-scale data centers;
Interface Compatibility: Standard PCIe interface, widely compatible with mainstream AI servers;
Open Ecosystem: Supports mainstream deep learning frameworks and provides in-depth customization through SDKs.
At the launch event, Suyuan Technology also released the "驭算" computing and programming platform. It supports mainstream deep learning frameworks, providing a complete compilation, debugging, and optimization toolchain. It also opens the SDK at the hardware level, providing deep developers with fine-grained computing programming interfaces. For complex model cluster training, such as BERT, it provides a distributed scheduling system optimized for ESL. With a 1024-node cluster scale, the training linear speedup ratio reaches 86%.
"Since its inception, Suyuan Technology's vision has been to provide customers with domestically produced general-purpose AI solutions for cloud training and inference, empowering service providers in various industries such as finance, healthcare, education, and transportation, and providing AI computing power for city clouds, government clouds, and public clouds." Zhao Lidong, CEO of Suyuan Technology, stated, "We have already established close cooperation with Tencent on projects related to general-purpose AI application scenarios, and we will expand to more AI application scenarios in the future. Starting with the new products released this time, Suyuan will provide more cost-effective, energy-efficient, and open-source solutions. We can also provide customized software and hardware support and services to meet the differentiated needs of various application scenarios, bringing greater value to our customers."
It is understood that Cloud燧T10 will be launched in the first quarter of 2020.
