NVIDIA Announces New AI Technologies at GTC 2020

Original article was published by Yisela Alvarez Trentini on Artificial Intelligence on Medium


NVIDIA’s GTC 2020 keynote presentation kicked off with NVIDIA’s CEO Jensen Huang making major announcements in data centers, edge AI, and collaboration tools. Among them is the video conferencing platform Maxine, the disease-fighting AI NVIDIA Clara Discovery, and the opening of beta programs for NVIDIA Jarvis and NVIDIA Merlin.

As the mobile cloud matures, the age of AI is beginning. Huang stated in the opening of his keynote that the powerful computation of graphics processing units ( GPUs) coming together are producing a big bang of modern AI. Automation lowers costs to consumers, makes life more convenient, and leads to breakthroughs.

NVIDIA is convinced that AI is changing everything, and the company wants to be precisely in the intersection that touches every facet of modern life. The GTC 2020 event was the largest so far, with over a thousand sessions in English, Chinese, Korean, Japanese, and Hebrew all in local times.

The GPU maker is worth approximately $251 million, a value that increased 70% over the past year. The expectations for the company are high, but NVIDIA is showing it’s definitely up for the challenge.

NVIDIA Maxine

NVIDIA Maxine is a cloud-native video streaming AI platform for video conferencing. With state-of-the-art AI models and optimized pipelines, Maxine can run several features in real-time in the cloud and reduce video bandwidth using AI video compression.

Maxine includes innovations such as face alignment, re-lighting, and gaze correction. Its face re-animation feature, for example, can identify key facial points of each participant on a video call and use these points to reanimate a person’s face so that people appear to be facing each other during a call.

The technology also supports super-resolution, which can convert lower resolutions to higher resolution videos in real-time, lowering the bandwidth requirement. It also offers translation, closed captioning, and noise removal. Using NVIDIA Jarvis, developers can also integrate virtual assistants that can take notes, set action items, and answer questions using human-like voices. And because all of these services run on the cloud, they can be offered to users on any device.

NVIDIA Clara Discovery

Understanding disease is one of humanity’s greatest endeavors, and drug discovery is particularly hard. It takes, on average, a decade and $2.5 million in R&D to develop a pharmaceutical drug, yet it fails in 90% of the cases. COVID-19 is an example of how important it can be to continue looking for more effective solutions to this challenge.

Using breakthroughs in computer science, simulation can now be used to understand the biological machinery that affects disease and search for candidates for new drugs. NVIDIA Clara Discovery is a GPU-powered drug discovery framework that combines the power of accelerated computing, AI, and machine learning.

Clara Discovery supports research in genomics, proteomics, microscopy, computational chemistry, virtual screening, visualization, and clinical imaging. The framework is optimized to run on NVIDIA DGX A100, the world’s most advanced AI system. With partnerships with Schrödinger, Structura, and Vyasa, this easy-to-deploy infrastructure can support the next generation of drug discovery.

NVIDIA Jarvis and Merlin Open Beta

With more people working from home and learning, shopping, and visiting their doctor remotely, increased strain has been placed on conversational AI services and recommendation systems. In line with this, NVIDIA has just announced that its Jarvis and Merlin systems have entered open beta.

Jarvis is a fully accelerated application framework to build enterprise-level multimodal conversational AI services. These services use sensor fusion (integrating video cameras and microphones), can run real-time on GPUs, reduce AI latency to half, and include tools for composing new models, transfer learning, and deploying. Early adopters of Jarvis include Voca, Kensho, and Square.

Merlin is an advertising recommendation acceleration framework that is expected to have a huge impact on online advertising. On some of the world’s largest ecommerce sites, recommender systems can account for 30% of revenue. A 1% improvement in these systems can translate into billions of dollars in revenue.

NVIDIA’s collection of libraries, which includes tools for deep-learning and prediction, are optimized to support hundreds of terabytes of data and be accessible through APIs. Businesses can already sign up for the NVIDIA Jarvis beta and the NVIDIA Merlin beta.

TensorRT 7.2

TensorRT 7.2 is the latest version of NVIDIA’s high-performance deep learning inference SDK. TensorRT applications can perform 40 times faster than CPU-only platforms, optimize neural network models in all major frameworks, and be deployed to hyper-scale data centers and product platforms.

The interface can be used for video streaming, speech recognition, fraud detection, recommendation, and natural language processing. The latest version has 30x faster AI-effects for super-resolution, noise removal, and virtual backgrounds, and can significantly reduce latency for real-time services.

Developers can import trained models into TensorRT from every deep-learning framework. The interface SDK optimizes the use of GPU memory and bandwidth, selects the best data layers and algorithms based on the target platform, and maximizes the memory footprint for tensors efficiently. The design is also scaled to process multiple input streams.