Pure Storage has opened its arms to white box storage vendors after deciding that servicing the world’s biggest AI rigs is beyond any one storage company.
The all-flash storage pioneer’s FlashBlade//EXA platform, which it unwrapped today, sees the vendor’s inhouse technology focused specifically on handling metadata in hyperscale AI setups.
More mundane storage will be handed off to vanilla kit – although those white boxes could come from Pure itself.
International CTO Alex McMullan said the strategy “reflects a orthogonal change…in terms of making effectively a storage supercomputer for the next iteration of scale.”
Hyperscalers’ AI ambitions are so far-reaching that “the current product simply wasn’t intended to grow to the same size. So we’ve taken that decision to go where we are at this point in time.”
The incredibly expensive, and power hungry, GPUs that hyperscalers and model developers are spending so many dollars on are already struggling to hit 25 percent utilization, he said, because legacy storage architectures weren’t up to the job. Meanwhile, those GPUs are still sucking in the equivalent power of a typical UK household.
“The hyperscalers are trying to save every single watt they can, because every watt they save, they can turn on another GPU,” he said.
These organizations are looking at 50TB per second or better data rates, he added. “We just don’t think that’s going to be doable with our existing architecture.”
Pure’s solution is to take its FlashBlade architecture and use that “as a metadata engine for the next generation storage supercomputer.” That layer will connect to the GPU cluster over NFSv4.1 over TCP. Meta data has emerged as a key bottleneck in AI systems.
Then, he said, “If we start with something that can run a SuperPOD on its own and just use that for metadata, then we should be able to keep up with everything else that sits around it.”
The everything else comes in the shape of software-defined data nodes connecting into the cluster via NFSv3 over RDMA. The metadata engine will then direct the GPU nodes to the data nodes layer.
These could be white boxes, he said, or in time a Pure Storage node. “White boxes just run on Linux kernel with an NFS server,” said McMullan. “They are standard. Vanilla.” The only requirement on the white boxes will be to sustain two 400 GB NICs.
The metadata engine could grow to “hundreds of petabytes if we want it to. Honestly, if the metadata is 100 petabytes, we have a bigger problem on its own.” As well as tackling the GPU utilization problem, the architecture will make it easier to have AI datasets in one place, reducing the impact of fragmentation.
McMullan said the ultimate target was the problems hyperscalers and similar operators were expecting to tackle in two years’ time. “Our existing flash blade is good enough for what’s there today, but some of the new requirements customers are talking about are making us take that pause.”
He said the company had a system in the lab running around 300 nodes delivering roughly 30 TBps.
Some early access beta customers were already in place, with “three or four of those in each of the theaters.” General availability for FlashBlade//EXA is slated for the summer. But Pure itself will have to join the queue for Nvidia certification, McMullan said, which should be signed off in the second half of the year.
McMullan was clear that the product was targeted at a small section of the market, comprised of “AI natives, people are building models, who are tuning models, who are looking at RAG-based capabilities.”
So, effectively hyperscalers and the biggest of AI factories. It could also include some life sciences organizations and some telco customers. And it will likely include some high end government customers.
It’s unlikely that many enterprises will be tyre-kicking these systems though, McMullan said.
“We don’t think there’s a play. Although, given all the things that have happened in the last couple of years in this space, who knows?”
