As human-caused climate change warms the planet, creating drier conditions across the Western U.S., wildfire intensity has grown. California’s wildfires over the last few years have devastated land, families, and communities. Rising tree and undergrowth density over the decades has provided fuel for catastrophic fires. Ongoing drought has created tinderbox conditions around the state and a near year-round fire season. These conditions are repeated in Oregon, Washington, Montana, Alaska, and Western Canada.

What’s on the ground matters. Crown fires, the devastating wildfires that spread embers and start spot fires, result from ladder fuels that help to create today’s mega fires, such as the Dixie fire, which burned nearly one million acres. Controlled burns help reduce this type of fuel load. But millions of acres of forest lands, comprising healthy forests, burned wastelands awaiting replanting, and recovering lands where growing non-conifers change the landscape and fuel load, present vast challenges for limited resources of management agencies. The challenges go beyond the forests to grasslands, scrubs, and other vegetation zones under threat of ignition by humans, which cause 95 percent of California’s fires and lightning (source).

For fire managers, fire scientists, and firefighters, analyzing and understanding fuel loads on both a macro- and microscale can help inform them about how to care for lands and more effectively mitigate and fight wildfires. Satellite and aerial imagery provide a useful source for analyzing what lies on the ground. Large, accumulative data from satellite and aircraft offer more insight and ground truth than small surveys by researchers in the field. There’s just too much ground to cover. Manually analyzing thousands of square meters of imagery takes an enormous amount of time.

“There is a lot of interest in using AI for wildfire management,” Mai Nguyen, a researcher and the Lead for Data Analytics at University of California San Diego’s (UC San Diego) San Diego Supercomputer Center (SDSC) commented. “UC San Diego and the University of California system as a whole are involved in this kind of research.”

Nguyen is developing tools to speed image analysis of fuel loads. She and other researchers are planning to use a new experimental supercomputer built specifically for AI acceleration. The system, called Voyager, is built on Habana Labs’ GAUDI training accelerator and GOYA inference accelerator, designed around their new Tensor Processor Core (TPC) architecture, and third-generation “Ice Lake” Intel Xeon Scalable processors. Each training and inference accelerator contains eight Tensor Processing Units (TPUs).

“The goal is to understand the land cover composition of an area in the context of wildfire management,” Nguyen stated. “Fire behavior depends on a lot of environmental conditions, and the fuel available to it is important. Grass creates one kind of fire, while dead wood presents another, potentially more intense fire.”

Nguyen has used AI for a number of applications. Besides image analysis, she has applied deep learning techniques to interdisciplinary problems that include disaster management and natural language processing. For image processing, she uses a pipeline she and her WIFIRE colleagues have applied to several diverse applications, including to understand the composition of a city, how quickly a refugee camp can build up, and to detect where schools are located in rural parts of Africa.

“We hope to combine our AI models with fire science models and fire science expertise and put together a platform that integrates all these different technologies for researchers to go to. They can study and simulate fire behavior under certain conditions to better understand how to fight wildfires.”

A New Architecture for Scientific Exploration

Nguyen has developed her Deep Learning (DL) algorithms on the TensorFlow framework for running on GPU-based systems. Voyager offers her an alternative architecture to test and develop her work with an easy transition to the new accelerators that use TPUs and CPUs instead of GPUs.

While GPUs traditionally have been the go-to architecture for large-scale deep learning training workloads, new technologies, like Habana Labs’ Gaudi and Goya accelerators, are emerging to give researchers new technologies at scale to explore. Voyager allows them to experiment, learn, and inform of new approaches that can address some of the most pressing research challenges. Funded by the National Science Foundation (NSF), Voyager is one of the first NSF AI-focused supercomputers built so data scientists can take advantage of this new architecture.

“We talked to several scientists about what they needed for their research,” Amit Majumdar, director of the Data Enabled Scientific Computing (DESC) division at SDSC said. “AI is a discipline itself and becoming an important component of their research. When the National Science Foundation requested proposals for unique experimental supercomputers, we began architecting Voyager and sought an NSF grant. This is hardware specially built for AI, both training and inference. We need this hardware to experiment, test, and learn in order to advance AI approaches.”

Voyager includes 42 training nodes of Supermicro X12 GAUDI Training Servers with Intel’s latest third-generation Xeon Scalable “Ice Lake” processors. Each training node contains eight Gaudi training cards. Two Supermicro SuperServer nodes are deployed for inferencing, with second-generation Intel Xeon Scalable “Cascade Lake” processors and eight Goya inferencing cards per node.

AI training on today’s massive datasets requires huge systems with a very large number of processor cores and nodes in order to train algorithms in reasonable time. At this scale, communication across the system is often constrained by the network. Each Gaudi training processor integrates ten 100 Gbps RDMA over Converged Ethernet (RoCE) interfaces. The interconnect can be configured several ways, making it flexible for different applications.

In Voyager, each of the eight Gaudi cards contained within the servers dedicates seven 100 Gbps ports to connect in an all-to-all, non-blocking configuration to the other cards. The other three 100 Gigabit ports are dedicated to scale out, giving each Voyager node 24 100 Gigabit ports. The scalability with integrated RoCE makes the overall system more efficient, according to Susan Lansing of Habana Labs, an Intel-owned company. Recently, Amazon Web Services added Gaudi cards to its EC2 instances as alternatives to their GPU instances.

For inference, Goya uses the same TPUs as Gaudi for accelerated inference at low power. Built on Habana’s Tensor processor core architecture with eight programmable cores in each inference card, Goya accelerates AI workloads, irrespective of the architecture on which they were trained. It natively supports many mixed-precision data types, including FP32, INT32/16/8, UINT32/16/8.

“We’re working with scientists who will run both training and inference,” Majumdar added. “Some will migrate their workloads built on other technologies to Voyager. Others will develop their models directly on Voyager. They will also need to transfer their models from training to inference, so it’s good to have both in one system.”

The NSF experimental program for Voyager allows access by a small group of scientists in close collaboration with SDSC and Habana application experts for three years. After three years of research, experimentation, development, and sharing their findings with their communities, Voyager will be available for general scientific research. Nguyen’s work is one of the many disciplines running these early experimental projects. Others include biology, genetics, materials science, atmospheric and astronomic sciences, and high-energy physics.

“We will be building and testing algorithms, optimizing them, and contributing to the community what we learn from Voyager and its technologies,” Majumdar said.

From GPUs to TPUs

Many data scientists create their algorithms to run on GPUs using DL frameworks, like PyTorch and TensorFlow. According to Lansing, the company’s software stacks and application suites are designed to ease development and building of new models or simplify the migration from existing GPU-based deep learning and inferencing architectures to Gaudi and Goya. The Habana Labs software suite is integrated with TensorFlow and PyTorch frameworks, and performance-optimized for Gaudi training and Goya inference deployment. While Voyager is being readied for its first three years of experimentation, SDSC has already had some access to a Gaudi-based test cluster at Habana Labs. That access has helped ease the migration to Voyager.

“Scientists are not doing anything special for Habana Labs technology,” commented Majumdar. “They are able to run their TensorFlow applications, essentially without any change. Their codes needed tweaking of only system-related parameters.”

Nguyen has begun running some of her codes on the Habana Labs cluster and expects her algorithms to easily migrate to Voyager.

“For the most part, it is straightforward to adapt my deep learning code to run on the Habana system,” she concluded. “This is important for adoption and ease of use for researchers like me.”

Voyager is getting a lot of interest around the scientific communities, according to Majumdar.

“Voyager is now deployed. We are doing the initial benchmarking. Many scientists are ready to work and lining up for time on the new system,” he concluded.

Ken Strandberg is a technical storyteller. He writes articles, white papers, seminars, web-based training, video and animation scripts, and technical marketing and interactive collateral for emerging technology companies, Fortune 100 enterprises, and multi-national corporations.