WekaIO, the leader in high performance, scalable file storage for data
intensive applications, today announced that its Matrix software
outperforms legacy file systems and local-drive NVMe on GPUs, delivering
an exceptional performance boost and cost-savings for high performance
AI applications when coupled with Mellanox Technologies' InfiniBand
intelligent interconnect solutions.
WekaIO Matrix achieved 5Gb/s of throughput per client running NVIDIA
TensorRT inference optimizer over Mellanox EDR 100Gb/s InfiniBand on 8
NVIDIA Tesla V100 GPUs, meeting the performance deep learning networks
require. Customers with demanding AI workloads could expect to achieve
11Gb/s 256K read performance from a single client host, a performance
benchmark that was achieved on a 6-host cluster with 12 Micron 9200
drives. Additionally, benchmark results showed that customers will
experience a boost in performance over local-drive NVMe even when link
speeds are reduced to 25Gb/s. Together, WekaIO with Mellanox InfiniBand
over local NVMe delivers performance superiority for customers who need
higher loads from single DGX-1 servers.
The work with Mellanox provides a comprehensive view and understanding
of WekaIO Matrix software's ability to distribute data across multiple
GPU nodes to achieve higher performance, scalability, lower latency, and
better cost savings for machine learning and technical compute workloads.
Mellanox is the leading supplier of performance interconnect solutions
for high performance GPU clusters used for deep learning workloads. When
customers couple Mellanox' ConnectX-series InfiniBand with WekaIO Matrix
software, they realize significant performance improvements to
data-hungry workloads without making any modifications to their existing
network.
"We are very excited by the results we are achieving with WekaIO," said
Gilad Shainer, VP of Marketing at Mellanox. "By taking full advantage of
our smart InfiniBand acceleration engines, WekaIO Matrix delivers world
leading storage performance for artificial intelligence applications."
"In deep learning environments we see large compute nodes, almost
universally augmented with GPUs, where customers need performance
scaling so that they can train their large neural networks faster. Local
file systems with NVMe fall short of the 5Gb/s of storage performance
required to leverage the processing power of GPUs-leaving expensive GPU
resources underutilized," said Liran Zvibel, co-founder and CEO at
WekaIO. "Our work with Mellanox demonstrates that WekaIO with InfiniBand
provides the best infrastructure for environments with GPUs, delivering
superior performance and economics for our customers."
See WekaIO Matrix demonstrated in Booth #624 at the NVIDIA GPU
Conference, March 27-29, 2018, in San Jose, Calif.