Jiwei network news, on March 21, Cambrian officially released the new training acceleration card MLU370-X8. MLU370-X8 equipped with dual-chip four-core Granny 370, integrated Cambrian MLU-Link multi-core interconnection technology, mainly for training tasks, in the industry widely used YOLov3, Transformer and other training tasks, the parallel performance of the 8-card computing system reached an average of 155% of 350W RTX GPU.
Cambrian Training Acceleration Card MLU370-X8
Dual-core Siyuan 370 architecture
According to Cambrian, the MLU370-X8 smart accelerator card provides a maximum training power consumption of 250W, which can take full advantage of the FP32, FP16 or BF16 computing performance commonly found in AI training acceleration. For the first time, Cambrian integrates the dual-chip four-core Granny 370 into the MLU370-X8 smart accelerator card, providing twice the memory, codec resources of the standard Siyuan 370 accelerator card, and equipped with MLU-Link multi-core interconnection technology. In the YOLOv3, Transformer, BERT, and ResNet101 training missions, the average performance of the 8-card parallel is 155% of the 350W RTX GPU.
Cambrian said that the MLU370-X8 smart acceleration card supports MLU-Link multi-core interconnection technology, providing in-card and inter-card interconnection functions. Cambrian specially designed MLU-Link bridge card for multi-card system, which can realize the full interconnection of 8 Siyuan 370 chips in a group of 4 acceleration cards, each acceleration card can obtain 200GB/s communication throughput performance, the bandwidth is 3.1 times that of PCIe 4.0, and multi-core multi-card training and distributed inference tasks can be performed efficiently.
According to the data, Cambricon NeuWare supports FP32, FP16 hybrid accuracy, BF16 hybrid accuracy and adaptive accuracy training and other training methods and provides flexible and efficient training tools, and the high-performance operator library has completely covered typical deep learning applications such as vision, speech, natural language processing, search recommendation and automatic driving, which can meet the needs of users for operator coverage and model accuracy.
Measured on the Cambriricon NeuWare SDK, on the common four deep learning network models, MLU370-X8 single card performance is comparable to the mainstream 350W RTX GPU; and in terms of multi-card acceleration, MLU370-X8 uses MLU-Link multi-core interconnection technology and Cambriricon NeuWareL communication library optimization to achieve a better parallel acceleration ratio in an 8-card environment.
MLU370-X8 complements the Siyuan 370 series product line
Cambrian has long adhered to the technical concept of "cloud edge and end integration, training and promotion integration, software and hardware collaboration". The MLU370-X8 provides twice the memory bandwidth of the Siyuan 370, combined with the MLUarch03 architecture and MLU-Link multi-core interconnection technology, the Advantages of the Siyuan 370 chip in the training task are fully utilized. The MLU370-X8 is positioned in the middle and high-end, combined with the high-end training products Siyuan 290 and Xuansi 1000, which further enriches the Cambrian training computing power delivery method, and cooperates with the MLU370-X4 and MLU370-S4 intelligent acceleration cards based on the Chiplet technology to form a complete cloud training and inference product portfolio.
The adaptation of the MLU370-X8 accelerator card with the domestic mainstream server partners has been completed and has been shipped to customers on a small scale.
Zhang Qiang, deputy general manager of Inspur Information Artificial Intelligence and High Performance Product Line, said: "Inspur and Cambrian are currently cooperating smoothly in the Siyuan 370 series of products, and have gradually landed in the fields of Internet, finance, manufacturing and other fields; the performance of MLU370-X8 is excellent, and we look forward to the two sides can continue to strengthen cooperation and bring excellent artificial intelligence computing power to more industries and customers." ”
Cambrian uses products to confirm its original intention and determination to customers: to provide excellent AI chip products for the explosion of artificial intelligence technology, so that machines can better understand and serve humans. (Proofreading/Arden)