Following an exciting Pure//Accelerate 2019 conference, I had the opportunity to capture an industry perspective from our guest speakers at the Global SI track.  Our first guest is Tony Paikeday, Director of AI Systems and data science platform at NVIDIA. Tony leads the product marketing team for the world’s first purpose built systems for enterprise artificial intelligence and deep learning.

Here’s what he had to say: 

Q:  What are the key industry trends that are interesting you at the moment?

A: We’re increasingly seeing more enterprises behave like supercomputing sites, spending massive capital on large scale training infrastructure. No longer exclusive to hyperscalers. They recognize that the most important business opportunities are tied to valuable use cases like NLP (chatbots, intuitive customer interactions, etc), and the modern NLP built on BERT is the way to go, but it needs massive compute power to train, using 8.3 billion parameters. What used to take a week to train last year, is now trainable in 58 minutes.

Q:  What are your clients asking for from you this year that they were not last year?

A: Last year we saw a lot of on-prem deployments from those who had the right data center facilities. Now we’re seeing a lot of clients who simply don’t have the desire to re-architect their data center for AI compute. So they’re asking for solutions to either host the infrastructure off-prem at colo, or rent PODS that are pre-staged at the colo.

Q:  Tell us about a big project that NVIDIA was involved in that you are proud of. What were the outcomes?

A: We’re really proud of the work being done at Paige.ai. They’re revolutionizing oncology and treatment of tumors with AI. AIRI built on DGX is at the center of their work, training their models on 25M pathology slides and images, and they just got FDA approval for their cancer detection process, just a year after starting work using AIRI.

Q:  Data and what businesses can do with it is clearly a big theme both in industry and at Accelerate. What leading indicators for change in the data space are you seeing in your business?

A: More customers are seeking ways to remove workflow bottlenecks at every step in the AI development workflow. We’ve obviously focused a LOT on training, but with RAPIDS for data ingest/manipulate and with what AI Hub is doing to unify formerly siloed storage architectures, we’re now seeing an end to end AI development pipeline that can truly mechanize or industrialize model development from concept to production.

Q:  If you were a strategist for an SI thinking about new Services offerings for next year, what would you invest in?

A: Definitely look at services that deliver AIRIaaS in partnership with leading edge colos. There’s a huge untapped market for customers who need massive training capacity, with deterministic and cost-effective performance, but in a utility model. The right offering paired with cloud, can ensure that customers get a low barrier to entry, and can then move model training to the facilities where their data already likely resides. It’s like offering them the best of both cloud and on-prem, just not on their prem.

Find out how Pure has made ‘AI First’ infrastructure a reality. 

0 Responses to NVIDIA Interview: Director of AI Systems and Data Science

Leave a Reply

Your email address will not be published. Required fields are marked *