At Oracle, customer service chatbots use conversational AI to respond to consumers with more speed and complexity.
Suhas Uliyar, vice president for product management for digital assistance and AI at Oracle, stopped by to talk to AI Podcast host Noah Kravitz about how the newest wave of conversational AI can keep up with the nuances of human conversation.
Many chatbots frustrate consumers because of their static nature. Asking a question or using the wrong keyword confuses the bot and prompts it to start over or make the wrong selection.
Uliyar says that Oracle’s digital assistant uses a sequence-to-sequence algorithm to understand the intricacies of human speech, and react to unexpected responses.
Their chatbots can “switch the context, keep the memory, give you the response and then you can carry on with the conversation that you had. That makes it natural, because we as humans fire off on different tangents at any given moment.”
Key Points From This Episode:
The contextual questions that often occur in normal conversation stump single-intent systems, but the most recent iteration is capable of answering simple questions quickly and remembering customers.
The next stage in conversational AI, Uliyar believes, will allow bots to learn about users in order to give them recommendations or take action for them.
Jared Ritter, the senior director of wireless engineering at Charter Communications, describes their innovative approach to data collection on customer feedback. Rather than retroactively accessing the data to fix problems, Charter uses AI to evaluate data constantly to predict issues and address them as early as possible.
What would the future of intelligent devices look like if we could bounce from using Amazon’s Alexa to order a new book to Google Assistant to schedule our next appointment, all in one conversation? Xuchen Yao, the founder of AI startup KITT.AI, discusses the toolkit that his company has created to achieve a “hands-free” experience.
Aakash Indurkha, head of machine learning projects at AI-based analytics platform Virtualitics, explains how the company is bringing creativity to data science using immersive visualization. Their software bridges the gap created by a lack of formal training to help inexperienced users identify anomalies on their own, and gives experts the technology to demonstrate their complex calculations.
Make Our Podcast Better
Have a few minutes to spare? Fill out this short listener survey. Your answers will help us make a better podcast.
NVIDIA and Microsoft are teaming up to provide the world’s most innovative young companies with access to their respective accelerator programs for AI startups.
Members of NVIDIA Inception and Microsoft for Startups can now receive all the benefits of both programs — including technology, training, go-to-market support and NVIDIA GPU credits in the Azure cloud — to continue growing and solving some of the world’s most complex problems.
The announcement was made at Slush, a startup event taking place this week in Helsinki.
With a variety of tools, technology and resources — including NVIDIA GPU cloud instances on Azure — AI startups can move into production and deployment faster.
NVIDIA and Microsoft will evaluate what startups in the joint program need, and how NVIDIA Inception and Microsoft for Startups can help them achieve their goals.
NVIDIA Inception members are eligible for the following benefits from Microsoft for Startups:
Free access to specific Microsoft technologies suited to every startup’s needs, including up to $120,000 in free credits in the Azure cloud
Go-to-market resources to help startups sell alongside Microsoft’s global sales channels
Microsoft for Startups members can access the following benefits from NVIDIA Inception:
Technology expertise on implementing GPU applications and hardware
Free access to NVIDIA Deep Learning Institute online courses, such as “Fundamentals of Deep Learning for Computer Vision” and “Accelerating Data Science”
Unlimited access to DevTalk, a forum for technical inquiries and community engagement
Microsoft for Startups is a global program designed to support startups as they create and expand their companies. Since its launch in 2018, thousands of startups have applied and are active in the program. Microsoft for Startups members are on course to drive $1 billion in pipeline opportunity by the end of 2020.
NVIDIA Inception is a virtual accelerator program that supports startups harnessing GPUs for AI and data science applications during critical stages of product development, prototyping and deployment. Since its launch in 2016, the program has expanded to over 5,000 companies.
Broadening support for GPU-accelerated supercomputing to a fast-growing new platform, NVIDIA founder and CEO Jensen Huang Monday introduced a reference design for building GPU-accelerated Arm servers, with wide industry backing.
Huang — speaking Monday at the SC19 supercomputing show in Denver — also announced that Microsoft has built NDv2, a “supersized instance” that’s the world’s largest GPU-accelerated cloud-based supercomputer — a supercomputer in the cloud — on its Azure cloud-computing platform.
He additionally unveiled NVIDIA Magnum IO, a suite of GPU-accelerated I/O and storage software to eliminate data transfer bottlenecks for AI, data science and HPC workloads.
In a two-hour talk, Huang wove together these announcements with an update on developments from around the industry, setting out a sweeping vision of how high performance computing is expanding out in all directions.
HPC Universe Expanding in All Directions
“The HPC universe is expanding in every single direction at the same time,” Huang told a standing-room only crowd of some 1,400 researchers and technologists at the start of the world’s biggest supercomputing event. “HPC is literally everywhere today. It’s in supercomputing centers, in the cloud, at the edge.”
Driving that expansion are factors such as streaming HPC from massive sensor arrays; using edge computing to do more sophisticated filtering; running HPC in the cloud; and using AI to accelerate HPC.
“All of these are undergoing tremendous change,” Huang said.
Putting an exclamation mark on his talk, Huang debuted the world’s largest interactive volume visualization: An effort with NASA to simulate a Mars landing in which a craft the size of a two-story condominium traveling at 12,000 miles an hour screeches safely to a halt in just seven minutes. And it sticks the landing.
Huang said the simulation enables 150 terabytes of data, equivalent to 125,000 DVDs, to be flown through at random access. “To do that, we’ll have a supercomputing analytics instrument that sits next to a supercomputer.”
Expanding the Universe for HPC
Kicking off his talk, Huang detailed how accelerated computing powers the work of today’s computational scientists, whom he calls the da Vincis of our time.
The first AI supercomputers already power scientific research into phenomena as diverse as fusion energy and gravitational waves, Huang explained.
Accelerated computing, meanwhile, powers exascale systems tackling some of the world’s most challenging problems.
They include efforts to identify extreme weather patterns at Lawrence Berkeley National Lab … Research into the genomics of opioid addiction at Oak Ridge National Laboratory … Nuclear waste remediation efforts led by LBNL, the Pacific Northwest National Lab and Brown University at the Hanford site … And cancer-detection research led by Oak Ridge National Laboratory and the State University of New York at Stony Brook.
At the same time, AI is being put to work across an ever-broader array of industries. Earlier this month, the U.S. Post Office, the world’s largest delivery service — which processes nearly 500 million pieces of mail a day — announced it’s adopting end-to-end AI technology from NVIDIA.
“It’s the perfect application for a streaming AI computer,” Huang said.
Next up for HPC: harnessing vast numbers of software-defined sensors to relay data to programmable edge computers, which in turn pass on the most interesting data to supercomputers able to wring insights out of oceans of real-time data.
Arm in Arm: GPU-Acceleration Speeds Emerging HPC Architecture
Monday’s news marks a milestone for the Arm community. The processor architecture — ubiquitous in smartphones and IoT devices — has long been the world’s most popular. Arm has more than 100 billion computing devices and will cross the trillion mark in the coming years, Huang predicted.
NVIDIA’s moving fast to bring HPC tools of all kinds to this thriving ecosystem.
“We’ve been working with the industry, all of you, and the industry has really been fantastic, everybody is jumping on,” Huang said, adding that 30 applications are already up and running. “This is going to be a great ecosystem — basically everything that runs in HPC should run on any CPU as well.”
World-leading supercomputing centers have already begun testing GPU-accelerated Arm-based computing systems, Huang said. This includes Oak Ridge and Sandia National Laboratories, in the United States; the University of Bristol, in the United Kingdom; and Riken, in Japan.
NVIDIA’s reference design for GPU-accelerated Arm servers — comprising both hardware and software building blocks — has already won support from key players in HPC and Arm ecosystems, Huang said.
In the Arm ecosystem, NVIDIA is teaming with Arm, Ampere, Fujitsu and Marvell. NVIDIA is also working with Cray, a Hewlett Packard Enterprise company, and HPE. A wide range of HPC software companies are already using NVIDIA CUDA-X libraries to bring their GPU-enabled management and monitoring tools to the Arm ecosystem.
The reference platform’s debut follows NVIDIA’s announcement earlier this year that it will bring its CUDA-X software platform to Arm. Fulfilling this promise, NVIDIA is previewing its Arm-compatible software developer kit — available for download now — consisting of NVIDIA CUDA-X libraries and development tools for accelerated computing.
Microsoft Brings GPU-Powered Supercomputer to Azure
“This puts a supercomputer in the hands of every scientist in the world,” Huang said he announced NDv2, a GPU-powered supercomputer now available on Microsoft Azure.
Giving HPC researchers and others instant access to unprecedented amounts of GPU computing power, Huang announced NDv2, a GPU-powered supercomputer now available on Microsoft Azure that ranks among the world’s fastest.
“Now you can open up an instance, you grab one of the stacks … in the container, you launch it, on Azure, and you’re doing science,” Huang said. “It’s really quite fantastic.”
For the first time, researchers and others can rent an entire AI supercomputer on demand, matching the capabilities of large-scale, on-premise supercomputers that can take months to deploy.
AI researchers needing fast solutions can quickly spin up multiple Azure NDv2 instances and train complex conversational AI models in just hours, Huang explained.
For example, Microsoft and NVIDIA engineers used 64 NDv2 instances on a pre-release version of the cluster to train BERT, a popular conversational AI model, in roughly three hours.
Magnum IO Software
Helping AI researchers and data scientists move data in minutes, rather than hours, Huang introduced the NVIDIA Magnum IO software suite.
A standing-room only crowd of some 1,400 researchers and technologists came to hear NVIDIA’s keynote at the start of SC19, the world’s top supercomputing event.
Delivering up to 20x faster data processing for multi-server, multi-GPU computing nodes, Mangum IO eliminates a key bottleneck faced by those carrying out complex financial analysis, climate modeling and other high-performance workloads.
“This is an area that is going to be rich with innovation, and we are going to be putting a lot of energy into helping you move information in and out of the system,” Huang said.
A key feature of Magnum IO is NVIDIA GPUDirect Storage, which provides a direct data path between GPU memory and storage, enabling data to bypass CPUs and travel unencumbered on “open highways” offered by GPUs, storage and networking devices.
NVIDIA developed Magnum in close collaboration with industry leaders in networking and storage, including DataDirect Networks, Excelero, IBM, Mellanox and WekaIO.
Academia, hyperscalers and scientific researchers have been big beneficiaries of high performance computing and AI infrastructure. Yet businesses have largely been on the outside looking in.
No longer. NVIDIA DGX SuperPOD provides businesses a proven design formula for building and running enterprise-grade AI infrastructure with extreme scale. The reference architecture gives businesses a prescription to follow to avoid exhaustive, protracted design and deployment cycles and capital budget overruns.
Today, at SC19, we’re taking DGX SuperPOD a step further. It’s available as a consumable solution that now integrates with the leading names in data center IT — including DDN, IBM, Mellanox and NetApp — and is fulfilled through a network of qualified resellers. We’re also working with ScaleMatrix to bring self-contained data centers in a cabinet to the enterprise.
The Rise of the Supercomputing Enterprise
AI is an accelerant for gaining competitive advantage. It can open new markets and even address a business’s existential threats. Formerly untrainable models for use cases like natural language processing become solvable with massive infrastructure scale.
But leading-edge AI demands leadership-class infrastructure — and DGX SuperPOD offers extreme-scale multi-node training of even the most complex models, like BERT for conversational AI.
It consolidates often siloed pockets of AI and machine learning development into a centralized shared infrastructure, bringing together data science talent so projects can quickly go from concept to production at scale.
And it maximizes resource efficiency, avoiding stranded, underutilized assets and increasing a business’s return on its infrastructure investments.
Data Center Leaders Support NVIDIA DGX SuperPOD
Several of our partners have completed the testing and validation of DGX SuperPOD in combination with their high-performance storage offerings and the Mellanox InfiniBand and Ethernet terabit-speed network fabric.
DGX SuperPOD with IBM Spectrum Storage
“Deploying faster with confidence is only one way our clients are realizing the benefits of the DGX SuperPOD reference architecture with IBM Storage,” said Douglas O’Flaherty, director of IBM Storage Product Marketing. “With comprehensive data pipeline support, they can start with an all-NVMe ESS 3000 flash solution and adapt quickly. With the software-defined flexibility of IBM Spectrum Scale, the DGX SuperPOD design easily scales, extends to public cloud, or integrates IBM Cloud Object Storage and IBM Spectrum Discover. Supported by the expertise of our business partners, we enhance data science productivity and organizational adoption of AI.”
DGX SuperPOD with DDN Storage
“Meeting the massive demands of emerging large-scale AI initiatives requires compute, networking and storage infrastructure that exceeds architectures historically available to most commercial organizations,” said James Coomer, senior vice president of products at DDN. “Through DDN’s extensive development work and testing with NVIDIA and their DGX SuperPOD, we have demonstrated that it is now possible to shorten supercomputing-like deployments from weeks to days and deliver infrastructure and capabilities that are also rock solid and easy to manage, monitor and support. When combined with DDN’s A3I data management solutions, NVIDIA DGX SuperPOD creates a real competitive advantage for customers looking to deploy AI at scale.”
DGX SuperPOD with NetApp
“Industries are gaining competitive advantage with high performance computing and AI infrastructure, but many are still hesitant to take the leap due to the time and cost of deployment,” said Robin Huber, vice president of E-Series at NetApp. “With the proven NVIDIA DGX SuperPOD design built on top of the award-winning NetApp EF600 all-flash array, customers can move past their hesitation and will be able to accelerate their time to value and insight while controlling their deployment costs.”
NVIDIA has built a global network of partners who’ve been qualified to sell and deploy DGX SuperPOD infrastructure:
In North America: Worldwide Technologies
In Europe, the Middle East and Africa: ATOS
In Asia: LTK and Azwell
In Japan: GDEP
To get started, read our solution brief and then reach out to your preferred DGX SuperPOD partner.
Scaling Supercomputing Infrastructure — Without a Data Center
Many organizations that need to scale supercomputing simply don’t have access to a data center that’s optimized for the unique demands of AI and HPC infrastructure. We’re partnering with ScaleMatrix, a DGX Ready Data Center Program partner, to bring self-contained data centers in a rack to the enterprise.
In addition to colocation services for DGX infrastructure, ScaleMatrix offers its Dynamic Density Control cabinet technology, which enables businesses to bypass the constraints of data center facilities. This lets enterprises deploy DGX POD and SuperPOD environments almost anywhere while delivering the power and technology of a state-of-the-art data center.
With self-contained cooling, fire suppression, various security options, shock mounting, extreme environment support and more, the DDC solution offered through our partner Microway removes the dependency on having a traditional data center for AI infrastructure.
There are a few different ways to think about pi. As apple, pumpkin and key lime … or as the different ways to represent the mathematical constant of ℼ, 3.14159, or, in binary, a long line of ones and zeroes.
An irrational number, pi has decimal digits that go on forever without repeating. So when doing calculations with pi, both humans and computers must pick how many decimal digits to include before truncating or rounding the number.
In grade school, one might do the math by hand, stopping at 3.14. A high schooler’s graphing calculator might go to 10 decimal places — using a higher level of detail to express the same number. In computer science, that’s called precision. Rather than decimals, it’s usually measured in bits, or binary digits.
For complex scientific simulations, developers have long relied on high-precision math to understand events like the Big Bang or to predict the interaction of millions of atoms.
Having more bits or decimal places to represent each number gives scientists the flexibility to represent a larger range of values, with room for a fluctuating number of digits on either side of the decimal point during the course of a computation. With this range, they can run precise calculations for the largest galaxies and the smallest particles.
But the higher precision level a machine uses, the more computational resources, data transfer and memory storage it requires. It costs more and it consumes more power.
Since not every workload requires high precision, AI and HPC researchers can benefit by mixing and matching different levels of precision. NVIDIA Tensor Core GPUs support multi- and mixed-precision techniques, allowing developers to optimize computational resources and speed up the training of AI applications and those apps’ inferencing capabilities.
Difference Between Single-Precision, Double-Precision and Half-Precision Floating-Point Format
The IEEE Standard for Floating-Point Arithmetic is the common convention for representing numbers in binary on computers. In double-precision format, each number takes up 64 bits. Single-precision format uses 32 bits, while half-precision is just 16 bits.
To see how this works, let’s return to pi. In traditional scientific notation, pi is written as 3.14 x 100. But computers store that information in binary as a floating-point, a series of ones and zeroes that represent a number and its corresponding exponent, in this case 1.1001001 x 21.
In single-precision, 32-bit format, one bit is used to tell whether the number is positive or negative. Eight bits are reserved for the exponent, which (because it’s binary) is 2 raised to some power. The remaining 23 bits are used to represent the digits that make up the number, called the significand.
Double precision instead reserves 11 bits for the exponent and 52 bits for the significand, dramatically expanding the range and size of numbers it can represent. Half precision takes an even smaller slice of the pie, with just five for bits for the exponent and 10 for the significand.
Here’s what pi looks like at each precision level:
Difference Between Multi-Precision and Mixed-Precision Computing
Multi-precision computing means using processors that are capable of calculating at different precisions — using double precision when needed, and relying on half- or single-precision arithmetic for other parts of the application.
Mixed-precision, also known as transprecision, computing instead uses different precision levels within a single operation to achieve computational efficiency without sacrificing accuracy.
In mixed precision, calculations start with half-precision values for rapid matrix math. But as the numbers are computed, the machine stores the result at a higher precision. For instance, if multiplying two 16-bit matrices together, the answer is 32 bits in size.
With this method, by the time the application gets to the end of a calculation, the accumulated answers are comparable in accuracy to running the whole thing in double-precision arithmetic.
This technique can accelerate traditional double-precision applications by up to 25x, while shrinking the memory, runtime and power consumption required to run them. It can be used for AI and simulation HPC workloads.
As mixed-precision arithmetic grew in popularity for modern supercomputing applications, HPC luminary Jack Dongarra outlined a new benchmark, HPL-ML, to estimate the performance of supercomputers on mixed-precision calculations. When NVIDIA ran HPL-ML computations in a test run on Summit, the fastest supercomputer in the world, the system achieved unprecedented performance levels of nearly 445 petaflops, almost 3x faster than its official performance on the TOP500 ranking of supercomputers.
How to Get Started with Mixed-Precision Computing
NVIDIA Volta and Turing GPUs feature Tensor Cores, which are built to simplify and accelerate multi- and mixed-precision computing. And with just a few lines of code, developers can enable the automatic mixed-precision feature in the TensorFlow, PyTorch and MXNet deep learning frameworks. The tool gives researchers speedups of up to 3x for AI training.
The NGC catalog of GPU-accelerated software also includes iterative refinement solver and cuTensor libraries that make it easy to deploy mixed-precision applications for HPC.
Researchers from the University of Tokyo, Oak Ridge National Laboratory and the Swiss National Supercomputing Center used AI and mixed-precision techniques for earthquake simulation. Using a 3D simulation of the city of Tokyo, the scientists modeled how a seismic wave would impact hard soil, soft soil, above-ground buildings, underground malls and subway systems. They achieved a 25x speedup with their new model, which ran on the Summit supercomputer and used a combination of double-, single- and half-precision calculations.
San Francisco-based Fathom, a member of the NVIDIA Inception virtual accelerator program, is using mixed-precision computing on NVIDIA V100 Tensor Core GPUs to speed up training of its deep learning algorithms, which automate medical coding. The startup works with many of the largest medical coding operations in the U.S., turning doctors’ typed notes into alphanumeric codes that represent every diagnosis and procedure insurance providers and patients are billed for.
Researchers at Oak Ridge National Laboratory were awarded the Gordon Bell prize for their groundbreaking work on opioid addiction, which leveraged mixed-precision techniques to achieve a peak throughput of 2.31 exaops. The research analyzes genetic variations within a population, identifying gene patterns that contribute to complex traits.
Nuclear Energy
Nuclear fusion reactions are highly unstable and tricky for scientists to sustain for more than a few seconds. Another team at Oak Ridge is simulating these reactions to give physicists more information about the variables at play within the reactor. Using mixed-precision capabilities of Tensor Core GPUs, the team was able to accelerate their simulations by 3.5x.
The Mile High City plays host next week to SC19, where GPUs will be key ingredients for computational science in some of the world’s most powerful supercomputers.
The race to AI and to exascale performance will be much of the buzz at the annual supercomputing event this year. For both, experts are relying on GPU accelerators.
In a special address Monday at 3pm MT, NVIDIA founder and Chief Executive Officer Jensen Huang will help kick off the conference. (Watch a mobile-friendly livestream here.) He’ll provide an in-depth look at the latest innovations in GPUs and how they’re transforming computational science and AI.
Modeling Brains, Earthquakes and More
A handful of demos at NVIDIA’s booth will give attendees a closeup look at how GPUs are pushing the envelope in science. NVIDIA Quadro RTX GPUs will host a visualization of an earthquake, and NVIDIA V100 Tensor Core GPUs will show a simulation of a human brain at nanometer-level resolution.
Ten partners will demo offerings using NVIDIA GPUs — ASRock Rack, Bright Computing, Boston, BOXX, Colfax, KISTI, Microway, One Stop Systems, Penguin Computing and Silicon Mechanics.
Huang’s overview is one of the first of many sessions on how GPUs can supercharge high performance computing with deep learning.
SC19 is host to three technical tracks, two panels and three invited talks that touch on AI or GPUs. For example, in one invited talk, a director from the Pacific Northwest National Laboratory will describe six top research directions to increase the impact of machine learning on scientific problems.
In another invited talk, the assistant director for AI at the White House Office of Science and Technology Policy will share the administration’s priorities in AI and HPC. She’ll detail the American AI Initiative the U.S. President announced in February.
Deep Dives in Deep Learning
A group of experts will give a deep dive Monday morning on how to tool high-performance computers for deep learning. They include senior engineers, scientists and researchers from Fraunhofer Institute, NVIDIA and Oak Ridge National Lab.
“Today we see excitement with machine learning being applied to many areas in computational science,” said Jack Dongarra, a professor at the University of Tennessee and one of three experts who maintain the TOP500 list of the world’s largest supercomputers. “As we go forward, I expect artificial intelligence to play an ever more important role in science.”
Back at NVIDIA’s in-booth theater, Marc Hamilton, vice president of solutions architecture and engineering, will kick off a slate of more than a dozen speakers, including talks from Mellanox on fast networking.
Other speakers will give updates on NVIDIA’s partnership to accelerate Arm-based supercomputers and on OpenACC, a parallel-programming model used on more than 200 applications. In a separate session Tuesday afternoon, Duncan Poole, the president of OpenACC, and a strategic partnership manager for NVIDIA, will host a birds-of-a-feather session on OpenACC.
Tracking the Race to Exascale
Meanwhile, many eyes are fixed on the exascale finish line for supercomputers able to calculate more than a quintillion floating-point operations per second or 1018 FLOPS. Getting to exascale, like breaking the petascale barrier in 2008, is a milestone in supercomputing that has recently galvanized the industry.
Arguably, the exascale era has already begun. Today’s most powerful supercomputer, the Summit system at Oak Ridge National Laboratory, has racked up a handful of exascale milestones. The 27,648 NVIDIA V100 Tensor Core GPUs in Summit can drive 3.3 exaflops of mixed-precision horsepower on AI tasks.
Harnessing some of that oomph, government and academic researchers shared the 2018 Gordon Bell Prize for using AI to determine the genetic roots of being susceptible to opioid addiction and chronic pain. Their work on one of America’s most pressing epidemics pushed the GPUs on Summit to 2.36 exaflops.
NVIDIA GPUs are now used in 125 of the TOP500 systems worldwide. Beyond Summit, they include the world’s second, sixth, eighth and 10th most muscular systems. Over the last several years, designers have increasingly relied on GPU accelerators to propel these big-iron beasts to new performance heights.
For more on NVIDIA events at SC19, check out our event page.
Italy energy company Eni is upgrading its supercomputer with another helping of NVIDIA GPUs aimed at making it the most powerful industrial system in the world.
The news comes a little more than two weeks before SC19, the annual supercomputing event in North America. Growing adoption of GPUs as accelerators for the world’s toughest high performance computing and AI jobs will be among the hot topics at the event.
The new Eni system, dubbed HPC5, will use 7,280 NVIDIA V100 GPUs capable of delivering 52 petaflops of peak double-precision floating point performance. That’s nearly triple the performance of its previous 18 petaflops system that used 3,200 NVIDIA P100 GPUs.
When HPC5 is deployed in early 2020, Eni will have at its disposal 70 petaflops including existing systems also installed in its Green Data Center in Ferrera Erbognone, outside of Milan. The figure would put it head and shoulders above any other industrial company on the current TOP500 list of the world’s most powerful computers.
The new system will consist of 1,820 Dell EMC PowerEdge C4140 servers, each with four NVIDIA V100 GPUs and two Intel CPUs. A Mellanox InfiniBand HDR network running at 200 Gb/s will link the servers.
Green Data Center Uses Solar Power
Eni will use its expanded computing muscle to gather and analyze data across its operations. It will enhance its monitoring of oil fields, subsurface imaging and reservoir simulation and accelerate R&D in non-fossil energy sources. The data center itself is designed to be energy efficient, powered in part by a nearby solar plant.
“Our investment to strengthen our supercomputer infrastructure and to develop proprietary technologies is a crucial part of the digital transformation of Eni,” said Chief Executive Officer Claudio Descalzi in a press statement. The new system’s advanced parallel architecture and hybrid programming model will allow Eni to process seismic imagery faster, using more sophisticated algorithms.
Eni was among the first movers to adopt GPUs as accelerators. NVIDIA GPUs are now used in 125 of the fastest systems worldwide, according to the latest TOP500 list. They include the world’s most powerful system, the Summit supercomputer, as well as four others in the top 10.
Over the last several years, designers have increasingly relied on NVIDIA GPU accelerators to propel these beasts to new performance heights.
The SC19 event will be host to three paper tracks, two panels and three invited talks that touch on AI or GPUs. In one invited talk, a director from the Pacific Northwest National Laboratory will describe six top research directions to increase the impact of machine learning on scientific problems.
In another, the assistant director for AI at the White House Office of Science and Technology Policy will share the administration’s priorities in AI and HPC. She’ll detail the American AI Initiative announced in February.
Radiology leaders have gathered for over 100 years running at RSNA, the annual meeting of the Radiological Society of North America, to discuss the industry’s latest challenges and opportunities. In recent years, AI in medical imaging has become a key focus — with startups at the center of the conversation.
Startups around the world are building AI solutions for a universal problem in medical imaging: limited time. Faced with rising numbers of patients being imaged, as well as the growing size of MRI and CT scans, radiologists must interpret one image every three or four seconds to keep up with the workload.
Agile startups are well-suited to tackle the demands of a rapidly evolving field like deep learning. In medical imaging, many are using AI to develop applications that target areas that slow radiologists down.
Healthcare startups raised more than $26 billion in venture capital funding last year and are partnering with major research institutions, hospitals and medical instrument manufacturers. They’re also receiving regulatory validation for clinical use: over three dozen healthcare AI startups have FDA clearance for algorithms that detect conditions including cancer, stroke and brain hemorrhages from medical scans.
At RSNA 2019, taking place in Chicago, Dec. 1-6, more than 50 attending startups are part of the NVIDIA Inception virtual accelerator program, which provides AI training and tools to fuel the growth of thousands of companies building GPU-powered applications, including over 700 healthcare startups.
Scan the Show for NVIDIA Inception Startups
Accelerated by NVIDIA GPUs, AI can speed up the acquisition, annotation and analysis of medical images to more quickly spot critical cases. It can also give experts quantitative insights that are too time-consuming to acquire using traditional methods.
Dozens of Inception companies will share their medical imaging applications for every phase of the radiology workflow at the RSNA AI Theater and the NVIDIA booth, including:
Higher-quality scans: Subtle Medical has developed the first and only AI software solutions FDA-cleared for medical imaging enhancement — SubtlePET for faster PET exams and SubtleMR for higher-quality MRI exams. Its software smoothly integrates with any scanner to enhance images during acquisition without altering the existing workflow, increasing efficiency and patient comfort. The company uses the NVIDIA DGX Station and NVIDIA DGX-1 to accelerate training, and NVIDIA T4 GPUs for inference.
Enabling AI-assisted annotation: TrainingData.io’s web platform helps researchers and companies manage their data labeling workflows, running on NVIDIA T4 GPUs for inference in Google Cloud. The startup leverages AI-assisted segmentation tools through the NVIDIA Clara Train SDK to label medical images that in turn train deep learning models for radiologists. And Palo Alto-based Fovia Ai, Inc. provides its customers with AI-assisted annotation powered by the NVIDIA Clara SDK in its tools for 2D and 3D visualization of medical images, which can seamlessly integrate into the clinical workflow.
Analyzing medical images: Tokyo startup LPIXEL develops deep learning image analysis tools using NVIDIA GPUs, including one to identify brain aneurysms from MRA, recently approved for clinical use in Japan. For lung tumor detection, China-based InferVISION’s AI tools identify and label lung nodules from CT scans in under 30 seconds. The company uses NVIDIA T4 GPUs for inference, achieving speedups of 4x over CPUs.
Processing surgical video: Doctors performing minimally invasive surgeries rely on live video feeds from tiny cameras to view the area they’re operating on. Kaliber Labs is building deep learning models that interpret these video feeds in real time for orthopedic surgery, identifying and measuring aspects of the patient’s anatomy and pathology, and providing intraoperative guidance to surgeons. The startup is using NVIDIA RTX GPUs for training and the NVIDIA Jetson AGX Xavier AI computing module for inference at the edge.
Rounding Out RSNA
In NVIDIA booth 10939 and beyond, we’ll be exhibiting the latest AI tools for medical imaging, from training to deployment.
In vitro fertilization, a common treatment for infertility, is a lengthy undertaking for prospective parents, involving ultrasounds, blood tests and injections of fertility medications. If the process doesn’t end up in a successful pregnancy — which is often the case — it can be a major emotional and financial blow.
Sydney-based healthcare startup Harrison.ai is using deep learning to improve the odds of success for thousands of IVF patients. Its AI model, IVY, is used by Virtus Health, a global provider of assisted reproductive services, to help doctors evaluate which embryo candidate has the best chance of implantation into the patient.
Founded by brothers Aengus and Dimitry Tran in 2017, Harrison.ai builds customized predictive algorithms that integrate into existing clinical workflows to inform critical healthcare decisions and improve patient outcomes.
Ten or more eggs can be harvested from a patient during a single cycle of IVF. The embryos are incubated in the lab for five days before the most promising candidate (or candidates) are implanted into the patient’s uterus. Yet, the success rate of implantation for five-day embryos is under 50 percent, and closer to 25 percent for women over the age of 40, according to the U.S. Centers for Disease Control and Prevention.
“In the past, people used to have to implant three or four embryos and hope one works,” said Aengus Tran, cofounder and medical AI director of Harrison.ai, a member of the NVIDIA Inception virtual accelerator program, which offers go-to-market support, expertise, and technology for AI startups revolutionizing industries. “But sometimes that works a little too well and patients end up with twins or triplets. It sounds cute, but it can be a dangerous pregnancy.”
The goal: a single embryo transfer that leads to a single successful pregnancy.
Going Frame by Frame
Embryologists manually analyze time-lapse videos of embryo growth to pick the highest-quality candidates. It’s a subjective process, with no universal grading system and low agreement between experts. And with five days of footage for every embryo, it’s nearly impossible for doctors to look at every frame.
Harrison.ai’s IVY deep learning model analyzes the full five-day video feed from an embryoscope, helping it surpass the performance of AI tools that provide insights based on still images.
“Most of the visual AI tools we see these days are image recognition,” said Aengus. “But with an early multi-cell embryo, the development process matters a lot more than how it looks at the end of five days. The critical event could have happened days before, and the damage already done.”
The company trained its deep learning models on a dataset from Virtus Health including more than 10,000 human embryos from eight IVF labs across four countries. Instead of annotating each video with detailed morphological features of the embryos, the team classified each embryo with a single label: positive or negative outcome. A positive outcome meant that a patient’s six-week ultrasound showed a fetus with a heartbeat — a key predictor of successful live births.
In a recent study, IVY was able to predict which embryos would develop a heartbeat with 93 percent accuracy. Aengus and Dimitry say the tool could help standardize embryo selection by reducing disagreement among human readers.
To keep up with Harrison.ai’s growing training datasets, the team upgraded their GPU clusters from four GeForce cards to the NVIDIA DGX Station, the world’s fastest workstation for deep learning. Training on the Tensor Core GPUs allowed them to leverage mixed-precision computing, shrinking their training time by 4x.
“It’s almost unreal to have that much power at your fingertips,” Aengus said. Using the DGX Station, Harrison.ai was able to boost productivity and improve their deep learning models by training with bigger datasets.
The company uses the deskside DGX Station for experimentation, research and development. For training their biggest datasets, they scale up to larger clusters of NVIDIA V100 GPUs in Amazon EC2 P3 cloud instances — relying on NGC containers to seamlessly shift their workflows from on-premises systems to the cloud.
IVY has been used in thousands of cases in Virtus Health clinics so far. Harrison.ai is also collaborating with Vitrolife, a major embryoscope manufacturer, to more smoothly integrate its neural networks into the clinical workflow.
While Harrison.ai’s first project is for IVF, the company is also developing tools for other healthcare applications.
Employers are scrambling to find people with AI, machine learning and data science skills and higher education is responding. Leaders from a group of top universities gathered at GTC DC Wednesday to discuss how universities can meet this demand.
Martial Hebert, dean of the School of Computer Science at Carnegie Mellon University, was joined by Cammy Abernathy, dean and professor of materials science and engineering at the University of Florida; Kenneth Ball, dean of the Volgenau School of Engineering at George Mason University; and Joe Paris, director for research computing at Northwestern University.
GTC DC has become the premier AI conference in the nation’s capital, this year attended by more than 3,600 developers, researchers, educators and CIOs focusing on the intersection of AI, policy and industry.
Wednesday’s panel, moderated by NVIDIA’s Jonathan Bentz, a solutions architect for higher education and research, life science, and high performance computing, discussed the importance of democratizing AI and data science tools and concepts for students.
The panelists explored three ways to better democratize AI: new degree programs, new coursework and building skills.
“We are distributing important digital skills throughout every course and major — from humanities to fine arts to healthcare to genomics — and developing brand new degrees to meet the needs of the changing workforce,” Ball said.
A major challenge to building skills, however, remains access to computing resources. Hebert described computing as “one of the biggest obstacles” faced by institutions limiting the number of students who can be involved with cutting-edge work.
In addition to access to the most capable machines, students need to be equipped with the knowledge and tools to address bias in AI.
“As we head down this path, it’s not lost on us the examples where our biases as programmers are finding their way into codes that are being applied to important tasks,” Paris said.
Abernathy said she’s “amazed” to see how quickly AI and machine learning have embedded themselves in almost every discipline. As the technology spreads, she stressed the importance of reaching out to and preparing underrepresented groups.
“It’s pretty clear if you want to be employable and a leader in your profession, you need to have skills in these domains,” Abernathy said. “It’s important that we provide access to a wider range of people.”
At GTC DC, the NVIDIA Deep Learning Institute offered a bevy of sold-out courses, workshops and hands-on training in AI, accelerated computing and data science and it announced a dozen new courses on Monday.
Resources:
Get started with your own AI training today with the DLI.
