putilov_denis - stock.adobe.com
Nvidia intros AI Omniverse Avatar generator for enterprises
Nvidia CEO Jensen Huang, in a conference keynote, promotes a host of new and updated tools and technologies, highlighted by an AI avatar maker for enterprises.
AI hardware and software giant Nvidia on Tuesday launched a new AI-based avatar generator for enterprises to use in augmented reality applications.
The vendor, at its GTC fall 2021 conference, also introduced a host of other new and updated AI technologies, including upgrades to its inference platform and tools to help enterprises train large language models.
"I think Omniverse is the most striking capability that an enterprise could bring on board," said Peter Rutten, an analyst at IDC.
Omniverse Avatar is a technology platform for generating AI avatars. Nvidia said the technology enables enterprises to create AI assistants that can be customized for any industry.
However, Rutten noted that the avatar platform will require a lot of compute power, but that should not dim enterprises' enthusiasm for it.
The Avatar is part of Nvidia Omniverse, a virtual world simulation and collaboration platform. During his keynote, CEO Jensen Huang showed various uses for the technology, including a customer-service avatar at a restaurant kiosk, a customer support avatar and a digital assistant.
Omniverse Avatar uses elements from speech AI, computer vision and natural language understanding.
"Nvidia's strategy of evolving the omniverse platform is uniquely differentiated," said Chirag Dekate, an analyst at Gartner. "From AI avatars to synthetic data generations … Omniverse is developing into a higher-level framework to create virtual worlds and enable digital twins use cases."
A digital twin is a virtual representation of a physical object or workflow.
Omniverse Avatar is currently in open beta.
Updating Triton Inference Server
Nvidia also updated its AI inference platform. The update includes a new version of Triton Inference Server and Nvidia TensorRT.
Peter RuttenAnalyst, IDC
The update to the Triton Inference Server is significant from a performance standpoint, Rutten said.
"This is important because AI inferencing is going to be a larger AI workload in the near future than AI training," Rutten said.
According to Dekate, the update shows how quickly the AI market is maturing.
"As more enterprises deploy and manage hundreds and thousands of models in production, accelerated infrastructures like Triton will be crucial in maximizing throughput, minimizing latency and accelerating deployment velocity," he said.
The new inference feature set includes new models and an analyzer to help enterprises understand how to optimize their AI models to run on a variety of different platforms.
The analyzer searches the areas between all the different GPUs and CPU options as well as all the different concurrency paths, and meets the required latency for picking the right model, said Ian Buck, general manager at Nvidia, during a media briefing.
"The right accuracy, the right amount of batching and the right hardware platform to get the maximum performance out of your infrastructure," he said.
The new version of Triton also supports multi-GPU multi-node inference and forced inference library, also known as RAPIDS FIL.
RAPIDS FIL is a new back end for GPU or CPU inferencing of decision tree models. Nvidia said that it provides developers with a unified deployment engine for deep learning and traditional machine learning.
Triton is available now from the Nvidia NGC catalog and as open source code from the Triton GitHub repository.
Nvidia also unveiled a new version of TensorRT that is now combined with TensorFlow and PyTorch. Nvidia said TensorRT 8.2 accelerates high-performance deep learning inferencing, high throughput and low latency in the cloud, on premises or at the edge. It's available now to members of the Nvidia developer program.
Partnership with Microsoft
During the virtual conference, Huang also revealed that Microsoft Teams is using Nvidia AI with Microsoft's Azure Cognitive Services for deploying AI to Teams. Nvidia AI will deliver live transcriptions and captions in 28 different languages, all running on the Triton inference server inside Azure Cognitive Services.
"Microsoft and all other cloud providers already partner with Nvidia to deliver GPU-accelerated training and inference services," Dekate said. "The Microsoft Teams announcement is about using Nvidia AI platforms to provide higher-accuracy AI services comprising extremely complex natural language models."
Partnership with Siemens Energy
Nvidia also touted its partnership with Siemens Energy.
The energy company is using Nvidia Triton for AI to help manage power plants and develop autonomous power plants.
"The Siemens/Nvidia combination solution has long ways to go," said Andy Thurai, analyst at Constellation Research. "But the promise of autonomous power plants is compelling."
Thurai said many checklists, compliance checks, inspection and approvals need to be done when a massive hydroelectric plant goes in and out of service. Since this requires a lot of workers and there's currently a skilled worker shortage combined with the older generation retiring, intelligent machines may be the answer to running power plants.
"Letting the machines learn instead of another generation of humans can be powerful," Thurai said. "The combination of video, audio, sensor inferences with millions of data points can be overwhelming for humans and even for normal machines to solve."
In his keynote, Huang also highlighted the new NeMo Megatron framework for training large language models, which now includes the Megatron 530B system that can be trained for new domains and languages.
Nvidia also unveiled during GTC Riva Custom Voice, a feature in the Riva speech AI system that enables enterprises to develop an expressive custom voice within hours using a small amount of data; and a new zero-trust platform that combines Nvidia BlueField DPUs, Nvidia DOCA kit (Data Center Infrastructure on a Chip), Nvidia Morpheus cybersecurity AI framework, and Nvidia Modulus, a framework for developing machine learning models.