AI Pilot to Production: Why Storage Matters, Per Pure Storage
At COLLIDE 2024 in Atlanta, Pure Storage's Chad Hendren said AI pilots often break when demand grows, because storage, metadata and power limits start to bite. He shared his company's claims on parallelized file tools, flash density and customer results.
By The AI Enterprise Magazine Team4 min read
Key takeaways
- Success creates the problem. A pilot people like quickly faces more users, new data and retraining demands.
- Feed the GPUs. Hendren said storage should keep GPUs busy and not starved for data.
- Metadata can dominate. He said a listing of about 7 million files fell from up to 15 minutes to about 30 seconds with parallelized tools.
- Power is a constraint. He said power saved at the storage tier can go to the GPUs.
- Predictable cost matters. Keeping data in place avoids egress charges and surprise bills, he said.
At COLLIDE 2024 in Atlanta, Chad Hendren, Consulting Field Solutions Architect at Pure Storage, argued that many AI pilots stall at the storage layer. In a session titled "Stack Your Enterprise's Odds for Success: AI Pilot to Production," he described what changes when a proof of concept becomes a popular application, and how his company said it addresses those demands. The conference was held October 10-11, 2024, at the Sandy Springs Performing Arts Center.
Hendren spoke as a vendor, and the performance figures below are his own claims from the session, not independent results.
Why do AI pilots break when they move to production?
Hendren described a familiar path. A company wants to "dip its toe in" AI, so it builds a simple model and proves the concept. It works, people like it, and the team pushes it out.
Then demand arrives. "Everyone's wanting to use it," he said, which puts new load on the infrastructure. Teams have to support multiple users, ingest new data, retrain the model so it stays accurate, and scale up without hitting roadblocks. A model only knows what was in its training data, he noted, so keeping it current is a continuing job.
He listed the pressures he sees in AI pipelines: storage and I/O-intensive training, metadata-heavy pipeline steps, siloed applications that force data to move between stages, power and energy budgets at petabyte scale, and the need for infrastructure that will not require a redesign as needs change.
What does storage have to do with GPU performance?
Hendren said the goal is simple: "keep those GPUs busy" and make sure they are not starved for data. He argued that storage that cannot keep up wastes the investment in expensive accelerators.
He also pointed to power. Pure Storage builds its own flash modules rather than using standard SSDs, which he said allows higher density at a power envelope up to 80% lower than competitors. He said the company was shipping a 150-terabyte module at the time, with a 300-terabyte version announced for the following year. Power saved at the storage tier, he said, can go to the GPUs.
On cost, he stressed predictability. "You want that predictable not just performance but also cost," he said, describing a cloud-adjacent model in which data stays in the customer's own storage while compute runs in the cloud, avoiding egress charges and surprise bills.
How much can metadata operations slow down AI workloads?
Hendren spent time on metadata, which he said can add significant time to training pipelines that handle millions or billions of small files. In a demo in his own lab, he said, listing a file set of about 7 million files (roughly 75 gigabytes) could take up to 15 minutes.
Pure Storage rewrote a set of standard Linux utilities to run in parallel on its scale-out file system. In his lab, he said, the listing step dropped to about 30 seconds, a 30x improvement for that one step. He said the toolkit is open source.
He also described a customer example. An oil and gas company moving data between Lafayette, Louisiana, and Houston was spending roughly two and a half hours on the transfer, and its deletes were so intensive that they halted all I/O on its existing platform. Hendren said the copy time fell to about three minutes and deletes became nearly instant, saving about five hours across those two steps.
What are real-world uses of AI image recognition?
Hendren noted that Pure Storage invested in Landing AI, a pre-IPO company working on large vision models, announced in June. He described three uses he has seen or knows of:
- A manufacturer with a robot-mounted camera that takes 15 to 20 pictures of each computer component at the end of the assembly line to catch defects such as missing connectors or a crooked CPU.
- A life sciences company using image recognition on medical scans to flag potential illnesses sooner than the human eye might.
- Domino's, which he believed put cameras over its oven exits in 2019 to check toppings and consistency, resulting in what he said was a 15% boost in quality ratings.
What about containers, tiering and reference architectures?
Because most modern analytics and AI tools ship in containers, which do not come with persistent storage by design, Hendren said Pure Storage's Portworx platform adds enterprise features such as backing up and cloning an entire application environment, on premises or across clouds.
He described a customer that removed a 400-server direct-attached application by separating storage from compute. He also described a tiering feature he called zero data movement tiering, and a cloud operating model for storage that lets a team with, for example, eight GPUs get a throughput-based SLA without a large upfront purchase.
Hendren said Pure Storage has had NVIDIA DGX BasePOD certification since about 2018 and was finishing certification for DGX SuperPOD, and he mentioned a Cisco validated design. For a retrieval-augmented generation system built with NVIDIA and presented at GTC, he said results were 36% faster than running the data on local SSDs, with a fraction of the energy.
Where to hear more from the data and AI community
COLLIDE 2027, the largest AI conference in the Southern U.S., will bring 1,000+ enterprise data and AI leaders to the Sandy Springs Performing Arts Center in Atlanta on Wednesday, October 13, 2027. Details are on the COLLIDE event page, and registration is coming soon.
Frequently asked questions
Why do AI pilots fail when moving to production?
Hendren said a successful pilot draws more users and new data, which strains infrastructure and requires retraining and scaling without roadblocks.
Why does storage matter for AI?
He said storage must keep GPUs busy and not starved for data, and must handle training I/O and metadata-heavy pipeline steps.
What did Pure Storage say about metadata performance?
In a lab demo, he said, listing about 7 million files fell from up to 15 minutes to about 30 seconds using parallelized Linux utilities.
How can AI image recognition be used in industry?
Hendren cited defect detection on an assembly line, medical scan review and pizza quality checks at Domino's.
This article is based on remarks by Chad Hendren, Consulting Field Solutions Architect, Pure Storage at COLLIDE Data + AI Conference 2024 on October 10, 2024. More coverage: COLLIDE AI Conference.
Filed under: Data & AI Infrastructure, AI Strategy & ROI
