NVIDIA GTC: Networking, Healthcare AI, Modeling and Simulation

By Allison Proffitt

November 9, 2021 | Jensen Huang’s kitchen—the site of the past several NVIDIA GTC Keynotes—deconstructed itself today in a slick simulated introduction as Huang, NVIDIA’s CEO, delivered his keynote highlighting the latest in AI and accelerated computing at NVIDIA.

Huang’s keynotes are always wide-ranging, touching on NVIDIA’s work across many industries including robotics, autonomous vehicles, an ambitious earth-wide digital twin for climate change, life sciences research, healthcare, cybersecurity, enterprise AI and more.

For life sciences AI/ML research, Huang focused on Python early on. He announced the latest accelerator for a Python libraries: cuNumeric accelerates NumPy with zero code change, scaling from one GPU to multiple GPUs to multi-node clusters. It joins other offerings in the NVIDIA RAPIDS open-source Python data science and machine learning ecosystem including cuDF, which is Pandas-like; cuML, which is Scikit-Learn-like; and cuGraph, which is NetworkX-like, Huang said. “cuNumeric is a data-center-scale math library,” he said.

In computing, Huang announced NVIDIA QUANTUM-2, what he called the most advanced end-to-end networking platform ever built. The 400 Gbps Infiniband platform consists of the Quantum 2 switch (five times the switching capacity of Quantum 1), ConnectX 7 NIC, the BlueField 3 DPU, and new software to enable performance isolation between jobs, congestion control, precision timing, multi-tenet bare metal security, and SHARP Gen 3 In-Network computing. It’s “the first networking platform to offer the performance of a supercomputer and the share-ability of cloud computing,” Huang said, a formerly impossible balance of performance and multi-tenant security. “With QUANTUM-2, your valuable supercomputer will be cloud-native and far better utilized.”

Focus on Healthcare and Life Sciences

In a pre-briefing last week, Kimberly Powell, Vice President of NVIDIA Healthcare, drilled into the astonishing growth in the market for healthcare data and outlined NVIDIA’s healthcare news.

“Healthcare data is 30% of all the world’s data, and by 2025 healthcare data will be growing at the highest compound annual growth rate of any industry at 36%,” Powell said. “The world’s medical AI development opportunity is tremendous, and it will remain so over the coming decades.” NVIDIA is committed to taking the state-of-the-art in computing architectures and applying it to this massive data opportunity, creating domain-specific tools and architectures for healthcare, she said.

She highlighted three results of these efforts—Clara Holoscan, Clara Monai, and Clara Discovery. The three products are bringing the latest in computing and applying it to healthcare, Powell said: “taking these transformer models that are used for NLP and helping to apply them in property prediction in molecules, or in protein structure prediction; taking all of the great things we’ve learned in self-driving cars and robotics and bringing it to the medical device community; taking everything from PyTorch and deep learning and bringing it into Clara Monai and surrounding it with all the tooling to make it healthcare-specific.”

NVIDIA Clara Monai for Healthcare AI

Calling it, “the world’s most advanced medical AI framework”, Powell introduced Clara Monai as the fruit of NVIDIA’s work with Kings College London and an open-source foundation of collaborators are building domain-specific tools and workflows. “Monai is the foundation core to the enterprise machine learning stack,” she said.

She listed the capabilities of Monai: AI data labeling, optimized training for super-fast experiment turnaround time, multi-modality training so we can build models that incorporate text with image data, auto ML training to automatically select network architectures and handle hyper parameter searches, and support for both batch and streaming application deployment to run research pipelines on real world data.

Since the launch of the framework last year, Powell says uptake has been astonishing with 150,000 download this year, averaging about 3,000 downloads per week. AWS and Azure both offer Clara quick-launch options and include Monai.

“Monai is dubbed the PyTorch of healthcare,” Powell said, “and the community of developers and partners is continuing to expand its reach.”

NVIDIA Clara Holoscan for Medical Devices

While medical data is growing at unprecedented rates, medical device innovation—everything from software to sensors to surgical robots—is exploding just as rapidly. Powell reports two million medical device types from 16,000 companies in over 10K modalities. And FDA approval for software as a medical device have grown 10-fold in 2 years, she said. She clarified that that isn’t all diagnostic technologies, but even algorithms to de-noise imaging, for example, are approved as software as a medical device.

“A medical device revolution is underway,” Powell said. “Sensor physics processing, data and image processing, and advance rendering in combination enabled by edge computing architectures that are remotely managed and upgradable are going to introduce a business model that is much like today’s self-driving car revolution. The business model will be revolutionized to enable medical device companies to evolve from hardware solutions into software-as-a-service solutions.”

Powell announced Clara Holoscan, an AI computing platform for scalable, software-defined, and end-to-end processing of data for devices. The platform seamlessly bridges devices with edge servers, she said, letting developers “compose containers and microservers into super fast pipelines.” Holoscan is enabled with NVIDIA Fleet Command so devices can be managed, orchestrated, and updated remotely.

Available today for Holoscan is the Clara AGX developer kit—"which is our first developer kit,” Powell interjected. The SDK will be available November 15. “The fact that we have all the NVIDIA engines from DPU, CPU, to GPU all in one common architecture and layering on top of that this ability to build streaming applications so you can have an end-to-end sensor-to-screen processing pipeline is exactly what the market needs,” she continued.

In the future, Powell said NVIDIA is building the orchestration layer to allow for microservices to be built into containers and those containers be linked together to create new, multimodality applications. For example, device companies are seeking to integrate voice for touchless use of devices. “You’re going to be able to piece together new workflows and new experiences at the edge for the clinical environment like we’ve never seen before.”

Powell also announced the Holoscan development computing platform consisting of the Orin 12 core ARM system on a chip and RTX A6000 Tensor Cores, CUDA cores, and RT Cores.

NVIDIA Clara Discovery for Faster Drug Discovery

“The future of drug discovery is computational—end to end—modeling the disease pathways, the genes involved, the drug-target interactions and off-target integrations,” both Powell and Huang said. “And with the confluence of AI for protein and chemical structure prediction and physics ML simulation approaches, we are witnessing the dawn of the biology revolution.”

Almost overnight the world went from having about 130 known [protein] structures to 130 million, she said, referring to the AlphaFold 2 and RosettaFold protein structure releases. In chemistry, Powell said, a similar explosion has happened for molecules structures.

NVIDIA is calling this the one million-X opportunity in drug discovery. “With this one million-x in possibilities, they must be explored in silico and through simulation. This has created a gigantic bottleneck in simulation.  New exploration in a physics and machine learning approach might come to the rescue.”

Powell highlighted Entos OrbNet, “a physics, machine-learned method to teach a graph neural network to replace the expensive quantum calculations,” she explained. OrbNet predicts Schrodinger’s equation, she said, delivering a thousand-fold increase in simulation performance. Powell demonstrated a simulation of a protein-candidate drug interaction. With OrbNet, the simulation took three hours on a GPU; without OrbNet it would have taken three months.

Clara Discovery comprises tools, models, and applications to enable in silico drug discovery including AI models like MegaMolBART and RoseTTAFold and simulation applications like Entos, Gromacs, and Torchani.