After a year of development, Huawei today announced the commercial availability of the Huawei Ascend 910 AI computing chip and the machine learning framework MindSpore. Huawei says the Ascend 910 is the world’s fastest AI processor, packing twice the performance of rival Nvidia’s Tesla v100.
Billed as the single chip with the greatest computing density, Ascend 910 delivers performance of up to 256 teraFLOPS under FP16 and 512 teraOPS under IN8 with declared max power consumption of 310W. In comparison, the GPU Tesla V100 delivers up to 125 teraFLOPS with a max power consumption of 300W, while Google’s TPU 2.0 with four ASICs can reach 180 teraFLOPS.
Ascend 910 is designed for deep learning training, and targets users such data scientists, researchers and engineers.
Huawei’s new ML framework MindSpore meanwhile provides device-edge-cloud training and inferencing based on a unified distributed architecture for machine learning, deep learning, and reinforcement learning, etc. It supports models trained on frameworks such as TensorFlow and PyTorch and provides flexible APIs decoupled from the core system.
Huawei Deputy Chairman and President of the company’s chip affiliate HiSilicon Eric Zhijun Xu said Huawei plans to open-source MindSpore in Q1 2020.
Huawei’s experiment results on ImageNet ResNet50 training demonstrate that the one-two punch of Ascend 910 and MindSpore is able to process twice as many images per second as a GPU and TensorFlow combo.
Huawei is doubling down on its hardware self-development efforts in a bid to offset damage from US government sanctions. Huawei was placed on the US Bureau of Industry and Security’s “Entity List” earlier this year, and in response the Chinese tech giant has taken steps to mitigate risks to its supply chain by introducing new homegrown AI chips, an AI framework, and the operating system HarmonyOS.
Xu also disclosed that Huawei is developing its own series of self-driving chips. The first, the Ascend 610, will be launched next year.
Journalist: Fangyu Cai | Editor: Michael Sarazen & Tony Peng