DDN works with NVIDIA

The NVIDIA Magnum IO software suite increases performance and reduces latency to move massive amounts of data from storage to GPUs in minutes, rather than hours. DDN testing has confirmed that, using the software’s NVIDIA GPUDirect Storage feature, the most intensive workflows will see significant improvements and benefit AI and HPC application output directly.  DDN expects to support the full NVIDIA Magnum IO suite, including GPUDirect Storage, in an EXAScaler^® EXA5 release in the middle of 2020.

 

“We value the deep engineering interlock between NVIDIA and DDN because of the direct benefits to our mutual customers,” said Sven Oehme, chief research officer, DDN. “Our companies share the desire to push the boundaries of I/O performance while simultaneously making deployment of these very large systems much easier.”

 

During testing with DGX SuperPOD, which itself is designed to deploy supercomputing-level compute very quickly, DDN was able to demonstrate that its data management appliance, the DDN AI400^TM, could be deployed within hours and a single appliance could support the data-hungry DGX SuperPOD by scaling as the number of GPUs scaled all the way to 80 nodes.  Benchmarks over a variety of different deep learning models with different I/O requirements representative of deep learning workloads showed that the DDN system could keep DGXSuperPOD system fully saturated.

 

“DGX SuperPOD was built to deliver the world’s fastest performance on the most complex AI workloads,” said Charlie Boyle, vice president and general manager, DGX Systems at NVIDIA. “With DDN and NVIDIA, customers now have a systemized solution that any organization can deploy in weeks.”

 

While the testing described above with DGX SuperPOD was performed with DDN’s AI400, DDN has since announced the AI400X^TM. The appliance has been updated to provide better IOPS and throughput and will ship with Mellanox HDR100 InfiniBand connections to support next-generation HDR fabrics. With these enhancements, the AI400X appliance could provide even better performance for AI and HPC applications.