Machine Learning Drives Change in Neural Network Development
Machine Learning, an important facet of artificial intelligence (AI), has been gaining prevalence in recent years due to its potential to do jobs less expensively and perhaps more efficiently than human workers. Computer scientists use a neural network, a computer system that obtains knowledge through learning, because it can do certain tasks sufficiently through training.
However, in the machine-learning development process, training is the most resource-intensive workload, according to Data Center Knowledge. In turn, deep-learning software development is creating a need for computer infrastructure that is not only specialized but also geared to handle specific workloads to train a neural network.
The demand for specialized computer infrastructure to train a neural network presents an opportunity for businesses. For instance, Cirrascale — a company based in Poway, California — switched from being a high-performance hardware vendor and a cloud-service provider to hosting and designing computer infrastructure for deep learning, according to Data Center Knowledge. Additionally, the company uses its data center to provide infrastructure-as-a-service in an approach that is fairly similar to Amazon Web Services (AWS), but with some noted differences.
“These types of boxes are very powerful,” David Driggers, the company’s CEO and founder, said in an interview with Data Center Knowledge. However, they also eat up a lot of power, and unlike AWS, Cirrascale is a “bare-metal” deep-learning cloud service; you get your box and use it to run your choice of software.
Cirrascale takes care of the setup, management and cooling of high-performance computing (HPC) clusters, an undertaking that appeals to customers new to HPC. Moreover, Cirrascale’s data center provides power densities above 30 kilowatts per rack, whereas a typical enterprise data center only has a power density of three to five kW per rack and rarely exceeds 10 kW, Data Center Knowledge reports.
“That’s a lot of wattage,” Driggers said. “Doing that part of it is hard, and we’re not charging a huge premium for that.”
For cooling, Cirrascale uses a proprietary liquid cooling system developed by ScaleMatrix. Furthermore, the company has many years of building HPC systems, according to the source.
“We’ve been doing 30 kW for well over 10 years, so we’re comfortable with standing up high-performance computing,” Driggers said.
However, it’s important to mention that Cirrascale’s previous experience with Verari Systems — an HPC hardware and data center container vendor — allows the company to practice critical skills gained, as well as put itself in the position to take on the opportunity of advancing neural networks. According to Driggers, because HPC environments and systems used to train neural networks are built using similar architectures, so when neural networks start to scale, they’ll look more like HPC systems.
Advancements and Outlook
Cirrascale is making strides to help advance neural network development. For instance, its top innovation — called PCIe Switch Riser — has a specific approach to interconnect GPUs in a single system, letting them talk to each other directly when bandwidth is high in order to improve performance and scalability, according to Data Center Knowledge.
However, unlike the innovations evident in AI systems like IBM’s Watson and Google’s AlphaGo, many companies are still in the development stages when it comes to using deep-learning technology in production. As a result, the best way to harness the technology and its future requirements is still uncertain, leaving plenty of room for companies to advance and train neural networks.