Pure Storage is elbowing into AI-based storage with FlashBlade, a use case that's a natural progression for the...
scale-out unstructured array.
The all-flash pioneer this week teamed with high-performance GPU specialist Nvidia to unveil Pure Storage AIRI, a preconfigured stack developed to accelerate data-intensive analytics at scale.
AIRI stands for AI-ready infrastructure. The product integrates a single 15-blade Pure Storage FlashBlade array fed by four Nvidia DGX-1 deep learning supercomputers. Connectivity comes from two remote direct memory access 100 Gigabit Ethernet switches from Arista Networks.
In this product iteration, Pure uses 15 midrange 17 TB FlashBlade NAND blades. Pure Storage claims a half rack of AIRI compute and storage is equivalent to about 50 standard data center racks.
The DGX-1 is outfitted with Nvidia Tesla V100 CPUs and rated to deliver 4 petaflops of performance. The Pure Storage FlashBlade's throughput is rated at 1.5 million IOPS.
Nvidia Cloud Deep Learning Stack embeds containerized versions for Apache MXNet, Caffe, Chainer, Python/Theano, PyTorch and TensorFlow AI frameworks. Pure developed code for the AIRI Scaling Toolkit software, which allows for rapidly testing and training analytic models simultaneously across multiple DGX nodes.
"Our AI-ready infrastructure came about like all good ideas: from our customers," said Matt Burr, Pure Storage vice president of sales for North America. "It was built to create an on-ramp for enterprises that wanted to implement AI but weren't sure how to buy the infrastructure."
Other storage vendors have integrated Nvidia technology into AI-related hardware. The Hewlett Packard Enterprise HPE Apollo 6500 Gen10 and IBM AC922 Power9 systems also can take advantage of Nvidia GPUs. High-performance computing vendor DataDirect Networks this week launched DDN ExaScaler DGX, based on its ES14K all-flash array integrated with the Nvidia DGX-1 deep learning system.
The Pure Storage FlashBlade scale-out NAS array uses NAND flash modules situated directly on the storage blades. FlashBlade supports file and object storage with the Purity operating system. Blades include 8 TB, 17 TB and 52 TB capacities. AIRI is preconfigured with 17 TB FlashBlade modules.
Storage can bottleneck GPU performance
Burr said Pure took the initiative to test a converged AI storage product after hearing from enterprises frustrated that their Nvidia processors often were idle due to disk-based storage.
"We bought a couple of DGX ourselves first to develop a testing toolkit and a testing configuration that could give linear scalability to go from a single DGX to multiple DGX units. When we asked Nvidia if this would be of value, the answer overwhelmingly was 'Yes,'" Burr said.
Pure customer Element AI switched to Pure Storage FlashBlade when its previous storage struggled to match the speed of Nvidia DGX nodes.
"We had settled on a low-end storage system, with plans to build a distributed storage architecture behind it. But what we discovered is that our GPU needs grew quickly beyond our ability to scale our storage. It got to the point where we were unable to effectively utilize our GPUs because of the bottleneck," said Jeremy Barnes, chief architect at Element AI.
"The Pure Storage FlashBlade is about 10 times faster out of the box, with no specific tuning or effort. It enabled us to boost our GPU from about 20% average utilization to close to 100% utilization."
Burr said pricing for Pure Storage AIRI is available from joint Pure and Nvidia resellers.