Suiyuan Technology wins the "China Chip • Annual Major Innovation Breakthrough Product" award for the second consecutive year
At the 16th "China Chip" Integrated Circuit Industry Promotion Conference held on December 20, Suiyuan Technology's second-generation AI training chip, "Suisi 2.0," won the highest award—"China Chip • Annual Major Innovation Breakthrough Product." Last year, the first-generation AI chip, "Suisi 1.0," also received this honor, making Suiyuan Technology the first company in the history of the "China Chip" selection to win this top award for two consecutive years with its AI chips.
Suiyuan Technology wins "China Chip • Annual Major Innovation Breakthrough Product" award for the second consecutive year
The "China Chip" selection event is one of the most influential and authoritative industry conferences in the domestic integrated circuit field. For 16 years, it has continuously encouraged product innovation, technological innovation, and application innovation in China's integrated circuit industry, promoting breakthroughs in key technologies by domestic integrated circuit companies and actively influencing and driving the overall development of China's semiconductor industry. This year's conference, themed "Chain up China Chip, Achieve China Made," received a record-high number of entries in the "China Chip" excellent product solicitation, with a total of 319 chip products from 217 chip companies.
Suiyuan Technology's second-generation AI training chip, "Suisi 2.0," released this July, underwent a complete upgrade and iteration. Compared to the first-generation chip, Suisi 2.0 increased FP32 computing power by 1.6 times, storage capacity by 4 times, storage bandwidth by 3.5 times, and interconnect bandwidth by 1.5 times, thus gaining the affirmation and recognition of professional judges and winning the "Annual Major Innovation Breakthrough Product" award again.
Suisi 2.0 achieved a leap in computing power through large-scale architectural upgrades. The new generation of fully self-developed GCU-CARA full-domain computing architecture supports comprehensive computing precision; the data flow computing acceleration engine provides computing capabilities during data flow transmission; the single-precision FP32 peak computing power reaches 40 TFLOPS, and the single-precision tensor TF32 peak computing power reaches 160 TFLOPS, both ranking first domestically.
At the same time, Suisi 2.0 is the first chip in China to support the world's most advanced HBM2E memory, significantly enhancing its training capabilities for ultra-large-scale models. It supports up to 64 GB of memory, and the peak storage bandwidth can reach 1.8 TB/s.
GCU-LARE full-domain interconnect technology is an interconnect technology developed by Suiyuan specifically for AI training clusters. It provides a bidirectional 300 GB/s interconnect bandwidth, supports interconnections of thousands of Yunsui T20 training accelerator cards, and achieves excellent linear scalability.
The independently developed full-stack software "Yusuan TopsRider 2.0," through software and hardware collaborative architecture design, can fully utilize the performance of Suisi 2.0; based on operator generalization technology and graph optimization strategies, it supports various model training under mainstream deep learning frameworks; open and upgradeable programming models and scalable operator interfaces provide customers with quick custom operator development capabilities for model optimization; fully optimized dynamic model support allows most models to seamlessly migrate to the Yusuan TopsRider 2.0 platform, making it developer-friendly; comprehensive compiler and toolchain support provides out-of-the-box graphical integrated development environment support, making deployment and maintenance more user-friendly for customers.
Suisi 2.0
Earlier this month, Suiyuan Technology launched a new generation of AI inference chips for the cloud, "Suisi 2.5," with a single-precision FP32 peak computing power of 32 TFLOPS and an integer INT8 peak computing power of 256 TOPS. Compared to the first-generation inference product, Suisi 2.5 increased floating-point computing power by 1.8 times and integer computing power by 3.6 times. It also innovatively uses high-bandwidth HBM2E storage, achieving an ultra-high bandwidth of up to 819 GB/s, providing high-throughput, low-latency performance for various cloud inference services, ensuring accurate, stable, and efficient operation. This also marks that Suiyuan Technology's complete product line in the cloud AI computing acceleration field, from training to inference, has been fully iterated to the second generation, enabling it to provide better and more complete AI computing solutions for customers in the Internet, smart city, smart government, and traditional industries such as finance, transportation, and energy with more competitive products.
"Suiyuan Technology's ability to win the 'China Chip • Annual Major Innovation Breakthrough Product' award again this year with Suisi 2.0 is both an affirmation and encouragement. The second-generation training and inference products, centered on the Suisi series of chips, have significantly improved performance indicators, and the upgraded Yusuan software platform is more convenient, user-friendly, and customer-friendly, providing customers with original innovation, high-performance, and cost-effective 'training + inference' complete solutions." Suiyuan Technology founder and COO Zhang Yalin said, "In the future, we will use three generations of chips in five years to narrow the gap with the world's leading products in the AI chip field. Suiyuan is executing with strong and consistent power according to plan, accelerating chip and product iteration. We will continue to adhere to the three business directions of the pan-Internet, traditional industries, and new infrastructure, serving the market and customers with new and more competitive products, and comprehensively and continuously creating value for customers."
