NVIDIA's AI push revolves around many fronts

NVIDIA has introduced new AI-related solutions recently to round out its AI value proposition. Project Holodeck, Drive Pegasus and Project Isaac are some recent announcements around the platform.

NVIDIA, known for its graphics acceleration technology in consumer gaming, is also going full speed ahead on using the same graphics acceleration knowhow to advance enterprise computing, including the field of artificial intelligence (AI).

Interest in AI has certainly grown in the last five years, says Dr David Kirk, NVIDIA Fellow and Former Chief Scientist at NVIDIA's recent AI Conference in Singapore. “The indicators are that we are entering into a new era of software and technology,” he said.

Indicators Dr Kirk listed include the number of graphics processing unit (GPU) developers growing by factor of 15 in last five years and the NVIDIA CUDA software development platform seeing downloads increase by a factor of five over the same period.

Additionally, US$6.6 billion has been disbursed for AI startup funding in 2017, a 12x jump in five years, while the number of deep learning papers published have grown 13x in three years. Three thousand papers were published from 2014 to 2017, he said. “For each paper there are another 100-200 researchers whose papers didn't get published or whose research isn't done yet,” he pointed out.

NVIDIA GPUs can be found in many applications and across many industries, from high performance computing to medical imaging, transportation, and logistics.

Interest and attendance at conferences has also grown, he said. The premier conference for AI, the Neural Information Processing Systems (NIPS) conference, reached 6,000 seats in 2017, and for the first time this year, reached full capacity more than a month before the beginning of the conference.

"It is not possible to get performance and throughput without the parallelism and power of GPUs," Dr Kirk said of deep learning, an AI technique.

GPU acceleration is pervasive in many industries, he said, listing applications in the high-performance computing, Internet services, transportation, medical imaging, and logistics sectors. "With deep learning and GPU acceleration, systems are able to do as well and sometimes even better than physicians," he said of the medical imaging field. Deep learning and GPUs can also solve large-scale planning problems like the routing of delivery trucks, organising products in a warehouse and efficient "last 50-mile" delivery for e-commerce, he added.

AI techniques such as deep learning are now "solving the unsolvable", Dr Kirk said. Companies are working with extremely complex tasks such as looking at partial images and inferring what the whole scene will look like, revolutionising the world of graphics. Path tracing is another challenge being tackled by AI. The technology traces photons as they bounce off objects and take on the colour of the object to increase the realism of graphics.

Other currently "unsolvable" challenges include robots which learn after being shown how to do something once; matching audio to facial animation realistically; using a video of a moving human body to animate a synthetic character; and understanding how an animated human character should move, so that it can automatically jump over obstacles. “It is very difficult to write software that actually works,” Dr Kirk said. “A neural network is able to learn how to do this very convincingly.”

NVIDIA is now looking ahead to how GPU technology can address future challenges with the NVIDIA AI platform, which addresses the need to train AI models to address a problem with frameworks, and then helps with the inferencing thereafter.

"Once we have the (neural) network built we must apply new data and new experiences to it to get something done," Dr Kirk explained. "There is an explosion in the number of intelligent devices and inferencing is the key to making that intelligence work."

The key is "more", Dr Kirk said. When there is more data, more computation power and more training, the AI solution is more accurate. "Bigger and faster networks get better results," he said.

The new NVIDIA Holodeck is part of the platform, and pitched as the design lab of the future.

Creation and collaboration in VR with the NVIDIA Holodeck.

Project Holodeck is a photorealistic, collaborative virtual reality (VR) environment that incorporates the feeling of real-world presence through sight, sound and haptics, was announced in May this year. It makes use of an enhanced version of Epic Games’ Unreal Engine 4 and includes NVIDIA GameWorks, VRWorks and DesignWorks. The environment allows creators - who can be from different locations - to import high-fidelity, full-resolution models into VR to collaborate and share with others, making time-to-decide faster.

"This is how products will be designed going forward," Dr Kirk said.

NVIDIA TensorRT is another component of the NVIDIA AI platform. The deep learning inference optimiser and runtime compiler delivers low-latency, high-throughput inferencing for deep learning applications. TensorRT can be used to optimise, validate, and deploy trained neural networks for inference whether it is a hyperscale data centre or an embedded solution.

Use cases include real-time services such as streaming video categorisation on the cloud, or object detection and segmentation on embedded and automotive platforms. TensorRT 3 will unlock optimal inference performance on NVIDIA's latest Volta GPUs. Performance comparisons include:

• Up to 40x higher throughput in under 7ms real-time latency versus using the CPU only for inference.
• Up to 3.7x faster inferences on the Tesla V100 compared to the Tesla P100 under 7ms real-time latency
• TensorFlow models optimised and deployed up to 18x faster compared to TensorFlow framework inferences on the Tesla V100

Source: NVIDIA website. TensorRT works better with NVIDIA technology such as the NVIDIA Tesla P100 data centre accelerator and the Tesla V100 data centre GPU than using the CPU only or a GPU with TensorFlow, the open-source library for machine intelligence. 

For self-driving vehicle research NVIDIA offers the DRIVE PX AI car computer platform to accelerate the production of automated and autonomous vehicles. The platform scales from a single processor configuration delivering AutoCruise capabilities to a combination of multiple processors and discrete GPUs designed to drive fully autonomous robotaxis. It already has 145 autonomous vehicle startups using it, Dr Kirk said.

Powerful systems are needed for self-driving, especially when there is more automation involved. "For a car to be fully safe, to be self-driving and reliable, it has to solve all the problems in multiple different ways, robustly," Dr Kirk said. 

Hamilton.

Marc Hamilton, VP of Solutions Architecture and Engineering, NVIDIA further explained that in levels 4 and 5 the vehicle must be able to sense the surroundings and understand where it is in the world. "It must have a high definition (HD) map, and use its perception in that map," he said, stressing that it is not a matter of orienting with a GPS system but true path planning, as GPS signals are not always available. "If it is travelling at 80kmh on a freeway and there is a slow car (ahead) it has to decide which way to pass."

At the higher levels of automated driving, the car has to understand where there is free space that it can occupy, and where is it not safe to drive, Hamilton said. Redundancy of both hardware and software is required for levels 4 and 5, he said, and even if companies have their own stack for autonomous vehicles, NVIDIA's stack can be used as a backup.

NVIDIA's self-driving robotaxi system, codenamed Pegasus, was launched October 11. It extends the NVIDIA DRIVE PX AI computing platform to handle level 5* driverless vehicles. NVIDIA DRIVE PX Pegasus delivers over 320 trillion operations per second (TOPS) - more than 10x the performance of its predecessor, the NVIDIA DRIVE PX 2.

For robotics, NVIDIA Isaac takes training virtual. The AI-based software platform lets developers train their virtual robot using detailed and realistic test scenarios. The simulations can be completed in minutes instead of months, all without the risk of damage to expensive robot components or injury to humans.

The company is also working with 1,900 startups through the Inception programme, which provides startups with the technologies and other resources they need. The startups are working  on diverse fields, such as healthcare, financial services, manufacturing, smart cities and drones.

"We are at the threshold of a new era in computing, of software that writes itself," Dr Kirk concluded. "What makes that possible are deep learning and GPU computing."

Explore:

Apply for early access to the Holodeck programme.

TensorRT 3 release candidate (RC) for the Tesla GPU (P4, P100, V100) and Jetson embedded platforms are available as free downloads to members of the NVIDIA developer programme. Download the RC.

Pegasus will be available to NVIDIA automotive partners in the second half of 2018. NVIDIA DriveWorks software and NVIDIA DRIVE PX 2 configurations are available today for developers working on autonomous vehicles and algorithms.

Explore:

Read the TechTrade Asia blog posts on:

How far AI can go: the Alibaba experience

Cutting-edge solutions from the NVIDIA AI Conference showfloor

Singapore's approach to AI

The EDB-NVIDIA Future Talents Program

The shared NVIDIA and NSCC platform for AI initiatives in Singapore

Hashtag: #NVAICONFERENCE17

*There are six levels of automation for driving, with level zero being no automation and level five, full automation.