On August 15, HKUST Xunfei held the "Xunfei Spark Cognitive Model V2.0 Upgrade Conference" in Hefei. At this press conference, iFLYTEK and Huawei jointly released the "Spark All-in-One Machine", an integrated software and hardware device for building exclusive large models.
On August 15, HKUST Xunfei held the "Xunfei Spark Cognitive Model V2.0 Upgrade Conference" in Hefei. At this press conference, iFLYTEK and Huawei jointly released the "Spark All-in-One Machine", an integrated software and hardware device for building exclusive large models.
Liu Qingfeng, chairman of HKUST Xunfei, said, "HKUST Xunfei and Huawei have jointly tackled the problem of computing power bottlenecks." In terms of single-card computing power, the "Spark All-in-One Machine" is benchmarked against
A100, and is also working with Huawei to build Domestic computing power clusters for ultra-large-scale large-scale model training form cluster advantages.
It is understood that in the research and development of large models, computing power and frameworks that can support parallel training of large-scale models with hundreds of billions of parameters are required, which usually requires thousands of high-power AI accelerator cards to run simultaneously and continuously. Based on this, the system architecture and the cluster capability of parallel training become the key to the development of large models.
It is understood that the Xinghuo all-in-one machine is based on Huawei Kunpeng CPU + Shengteng GPU, and also uses Huawei storage and network to provide the whole cabinet solution. FP16 (half-precision floating-point number format) has a computing power of 2.5 petaFLOPS (1 petaFLOPS is 10 million per second. million floating-point operations). The most popular Nvidia DGX A100 in large model training, its single node contains 8 GPUs, which can output FP16 computing power of 5 PetaFLOPS.
(The Ascend chip model used in the Spark all-in-one machine is suspected to be the Ascend 910B, and its performance has improved compared with the previous generation. The official website data shows that the FP16 computing power of the Ascend 910 is 320TFLOPS.)
According to reports, the integrated deployment of training and reasoning of the Xinghuo all-in-one machine can be used for applications in multiple fields such as question answering systems, dialogue generation, knowledge graph construction, intelligent recommendation, etc. It has large model pre-training, multi-modal understanding and generation, and multi-task Ability to learn and transfer.
The Xinghuo all-in-one machine also customizes the hardware for the training algorithm and reasoning application of the Xinghuo cognitive intelligent large model, which can greatly reduce the cost of use for enterprises. It can directly provide 5 customized optimization modes such as dialogue development, task arrangement, plug-in execution, knowledge access, and prompt engineering, as well as more than 10 out-of-the-box rich scene packages such as office, code, customer service, operation and maintenance, marketing, and procurement. , supports 3 model sizes for users to choose.
It is worth noting that last month, HKUST Xunfei announced that Xunfei Xinghuo will cooperate with Shengteng AI to create a new universal intelligent base based on China's independent innovation. On the one hand, Xunfei Xinghuo's cognitive large model is based on the integrated design of training and reasoning, which realizes the technological breakthrough of sparse large-scale models and low-precision quantization, and can efficiently adapt to Ascend AI, accelerating the industry's application and iteration of large models; On the one hand, with Ascend AI as the core, software and hardware are collaboratively optimized to build a large-scale model training cluster with centralized computing power, superior performance, stable supply, and data security.
In the speech "Xunfei Xunhuo Cognitive Large Model V2.0 Upgrade Conference" that day, Liu Qingfeng explained in more detail, "Huawei and HKUST Xunfei have jointly developed software and hardware platforms and software support tools for Ascend AI. , combining high-computing AI chips, high-performance operator libraries, multi-card high-speed interconnection, and distributed storage, especially our joint screening and polishing of the most important operator libraries required by artificial intelligence. Then here On the basis of the framework, iFLYTEK’s training and data closed-loop full-process design, as well as a self-developed large-model training platform designed for training and reasoning integration, supports large-scale heterogeneous computing power compatibility, and also supports hybrid cloud architecture for easy expansion , so that the demonstration of Xunfei Xinghuo V2.0 and all products we have seen today are built on a safe and controllable platform.”
