The annual NVIDIA GTA Conference opened in Beijing on September 26th. The much-anticipated GPU developers’ conference showcased the company’s latest endeavours in AI, deep learning, healthcare, VR, and self-driving cars.
In addition to deep learning engine TensorRT 3.0, NVIDIA introduced the HGX-1 hyperscale GPU accelerator powered by Tesla V100 for AI cloud computing, and the all-new AI processor Xavier for self-driving cars.
“Now IFlytech, WeChat, JD.com , Alibaba and other Chinese companies are using our products to expedite their products.” said company CEO Jensen Huang.
TensorRT 3.0 Deep Learning Engine
Jensen Huang spoke glowingly of TensorRT 3.0’s functionalities. TensorRT is a bridge between neural network frameworks and hardware (GPU) platforms, comprising a variety of chips that range in application from servers to individual devices. The chips can reduce neural network calculation latency, which is crucial for consumer-facing applications.
The new TensorRT 3.0 supports all kinds of popular neural network frameworks (including TensorFlow, Microsoft Cognitive Tookit, MXNet, PyTorch, Caffe2, PaddlePaddle, and the late Theano) and covers more GPU types (including the recently launched Jetson TX2 and Tesla V100) than its previous version. Based on image processing capabilities, the new TensorRT also supports more types of neural networks. As a programmable platform, TensorRT gives GPU advantages over other hardware options.
Introducing the HGX-1 Accelerator to China
HGX-1 hyperscale GPU powered by Tesla V100 was the second highlight of the day. In the morning of the General Assembly, Huang announced that NVIDIA is working with Huawei, Inspur, Lenovo and other companies to form partnerships around this product line.
The accelerator contains eight Tesla V100 chips (plus the CPU and other chips). NVIDIA claims performance is equivalent to 150 CPUs combined (for speech recognition, image recognition and other tasks). One HGX-1 can provide users with significantly increased CPU and GPU computing capabilities, and perform up to 100 faster than CPU-based servers in deep learning tasks, reducing training costs by 80%.
The World’s First AI Machine Processor: Xavier
NVIDIA believes that AI technology, user application (related to data), and computing power gave birth to the first batch of AI companies. As a chip supplier, NVIDIA hopes to cooperate with companies committed to further growing AI.
Ambitious e-commerce platform JD.com chose NVIDIA to power its UAV and robot applications. The company predicts that in 2022 it will deploy one million unmanned aerial vehicles for logistics.
NVIDIA’s solution is Xavier, a new generation of terminal Volta chips which will be added to the next generation of Jetson systems in 2018, providing support for JD.com’s logistics.
Journalist: Zenan Li | Localization: Meghan Han | Editor: Michael Sarazen