At this year’s GPU Technology Conference (GTC) in San Jose, NVIDIA announced its second-generation artificial intelligence (AI) system with 10x the performance of the first generation. This breakthrough in GPU performance opens up a whole new range of use cases in the deep learning and data science worlds, and it also underscores the critical need for a data pipeline that can keep up with the performance advances in GPUs.
Any team deploying NVIDIA’s DGX-2 has to ask themselves what it will take to feed data to a deep learning cluster with so much additional computing power. The impacts on networking, I/O bandwidth, I/O speeds, and data feeds, and the amount of fan-in required, are profound. NetApp is working to create reference architectures to deliver the extreme I/O performance required by next generation GPU’s.
Is AI Entering the Mainstream?
GTC is one of the premier conferences for AI and deep learning. The conference covers all aspects of GPU technology, including advanced graphics and GPU computing. Attendance has risen 10-fold over the years, spurred by interest in AI. This year’s conference confirmed NetApp’s contention that, as organizations operationalize AI, they need a data pipeline that offers an edge-to-core-to-cloud architecture.
I’ve been to a lot of AI technology conferences over the past 15 months, and at GTC I saw a significant shift that suggests that many organizations are moving beyond research and pilot projects. The increased participation of infrastructure architects at this year’s event is a good indication that companies are actively putting AI into operation. The high levels of interest kept those of us working the NetApp booth hopping, and our AI solutions clearly resonated with the attendees.
- Edge data solutions. Many attendees recognized the value of a solution that can connect the edge from both infrastructure and data management perspectives. As I described in a previous blog post, there are different I/O characteristics for data flowing in from the edge into a training cluster. A lot of the current edge solutions focus on the needs of specific vertical markets and combine compute and storage in a small footprint. By separating out storage, NetApp offers greater functionality in a solution suited to a wide range of markets. NetApp has partnered with Vector Data to deliver ruggedized solutions for edge locations that need them.
- Storage near the cloud. I had a number of conversations with people who saw the value of storing data near the cloud where it can be accessed with high performance from multiple public clouds. This allows organizations to take advantage of different public clouds as AI technologies evolve without getting locked in.
GTC also featured a variety of educational sessions, including a large number on implementing the infrastructure necessary for AI. These sessions underscored the need for flexible building blocks that can scale out as an AI installation grows. One thing that was a bit surprising to me was that the implementations discussed were almost entirely on premises.
A survey that NetApp conducted at GTC seemed to bear out the conclusion that serious AI practitioners see a need for on-premises AI infrastructure versus cloud. Although it’s important not to over-interpret the data, of 105 people we surveyed at the conference, about 60% were deploying AI on premises versus about 23% in the cloud.
Another trend that was apparent from the survey is the popularity of NFS as a filesystem for AI deployments, versus Hadoop, S3, and so on. (I discussed the pros and cons of various options in a previous blog post on choosing an optimal file system and data architecture.)
It’s also clear that the idea of a data pipeline that extends from edge to core to cloud resonated with respondents, with the largest number of respondents choosing that approach. NetApp was the only infrastructure vendor at GTC presenting solutions that cover the full range of data management needs across a data pipeline.
Overall, survey participants felt that the survey questions helped them sharpen their thinking about AI infrastructure planning and found it quite valuable. Look for NetApp to share complete survey results in the near future.
Just so you don’t think it was all work and no play at GTC, we had a number of fun demos. The “Vincent” demo system was developed by Cambridge Consultants (no connection to Cambridge Analytica). It takes your doodles and uses AI to render them in the style of Vincent Van Gogh, and some of the results are quite striking. The model was trained using NVIDIA and NetApp infrastructure.
We also featured an autonomous vehicle video animation that shows how NetApp technology can help design a solution to accelerate deep learning for self-driving.
Enterprises are moving beyond the investigation phase and starting to make AI operational. As that happens, there’s a growing need to move data quickly and efficiently. Given the rapid advances in GPU technology, data pipeline architecture is already on the critical path for AI.
We think that’s why NetApp’s edge-to-core-to-cloud message resonated with attendees at GTC. NetApp was the only infrastructure vendor at the show capable of supporting not just artificial intelligence, deep learning, and machine learning workloads but all the workloads that were discussed at the show, including graphics acceleration, HPC, cloud, database, and ERP—and all using a single platform: All Flash FAS.
Previous blog posts in this series: