Today the GPU Technology conference in Munich kicked off with a keynote by NVIDIA CEO Jensen Huang. NVIDIA announced the NVIDIA Holodeck, the Tensor RT 3 library, NVIDIA's Drive platform, and the Pegasus computer for autonomous taxis.
NVIDIA Holodeck
NVIDIA announced a new CAD editor in which you edit models in a virtual reality environment: the Holodeck. You can take your original CAD design and import it into this design tool. The holodeck renders it photo-realistically, and you can have a physically simulated interaction with your model. The graphics run at 90 frames per second giving you a feeling of directly working on your design. Eventually, an AI agent will be added to the Holodeck that can help you with your designs. An example feature of is AI is completing your model using only components that are already available in your component library. Another useful tool is the geometry clipping tool, which allows you to look inside a design you were building. An early access program is available at nvidia.com/holodeck.
Cloud service providers
Many cloud providers are now offering the NVIDIA Voltage GPU as part of their cloud services. Companies like Microsoft Azure, Google Cloud Platform, and Alibaba Cloud are now offering this new technology in the cloud. This allows startups to rent a GPU instead of having to build a supercomputer themselves.
TensorRT 3: Processing 5700 images per second
Jensen Huang showed how the TensorRT 3 library can speed up inference of your neural network. This library optimizes the inference of neural networks for several NVIDIA devices: from Tesla V100, Jetson, Drive PX2, Tesla P4. During inference, a neural network makes a prediction based on an input: the network is not trained in this phase. Inference is the phase in which you deploy a neural network in your service. TensorRT can infer faster by combining layers and Tensors, and the library traces paths through a neural network to find out what paths are most important, and what paths can be removed from your model. The result is that it can run inference on 5700 images per second, compared to only 300 using only TensorFlow and 140 per second without GPU acceleration. This comparison was made on a 50 layer ResNet: a deep neural network for image recognition with 50 hidden layers.
Autonomous vehicle efforts
NVIDIA announced several features for the autonomous driving industry. Their Drive platform consists of APIs that range from low-level to high-level features. Products in the Drive platform are the operating system Drive OS, the Drive PX computer, and the Driveworks SDK with algorithms for autonomous vehicles. Besides the already available Drive PX, they announced the Pegasus. Pegasus is a supercomputer designed for production and deployment of robotic taxis. The Pegasus has a 320 Teraflops processor, comparable with 400 CPUs. Although it has a lot of computation power it uses 500 watts, which translates to a longer range for electric robotic taxis.
NVIDIA also announced the Drive IX(Intelligent eXperience) SDK. It senses what is happening inside the vehicle: where your eyes gaze, your head pose, and voice recognition. Useful applications are alerting the user when he misses a dangerous situation and keeps the user alert. This announcement supports their safety approach: by solving the same problem in multiple ways NVIDIA wants to guarantee that even when one of their systems fails the car keeps driving safely.The Drive IX SDK will be available in the fourth quarter of 2017.