The two suppliers have both designed tech using Nvidia DGX-1 GPU servers and their own storage arrays to provide storage-server systems for "AI" applications such as deep learning.
Pure's deliverable hardware-software system is called AIRI (AI-Ready Infrastructure), while NetApp and Nvidia have deep learning reference architecture (DL RA). NetApp published Resnet-152 and Resnet-50 performance numbers and bar charts with 1, 2, 4 and 8 GPUs for its A700 all-flash array and one DGX-1, not the top line A800 array.
Pure contacted us and supplied the numbers for its Resnet runs, allowing a direct comparison. Resnet-152 first:
|Resnet-152||1 GPU||2 GPU||4 GPU||8 GPU|
The batch size is 64. Resnet-50 with the same batch size next:
|Resnet-50||1 GPU||2 GPU||4 GPU||8 GPU|
Pure's AIRI beats the NetApp/Nvidia DL RA at all GPU levels. Charting the numbers makes it plain:
Pure AIRI vs NetApp/Nvidia DL RA
AI system choices are going to need more than two simple benchmark runs, and the Resnet benchmarks have been described by a Pure spokesperson as "incredibly complex".
But here we do see a direct comparison between Pure's FlashBlade and NetApp's A700 – and the FlashBlade is ahead. ®