DDN asserts that the ideal storage solution for generative AI and other AI workloads requires a balance between read and write speed. DDN claims that their own storage system offers the best balance due to its superior write speed compared to competing systems. One of their offerings, the Exascaler AI400X2 array with Magnum I/O GPUDirect certification, is designed to work with Nvidia's DGX SuperPod AI processing system. This array utilizes 60TB QLC SSDs (4bits/cell) and includes a compression feature to enhance effective capacity. According to DDN, there are currently 48 AI400X2 arrays being used with Nvidia's largest SuperPODs, and the company claims to have shipped more AI storage in the first quarter of this year than in all of 2022. In a blog post titled "Exascale?Let's Talk, " James Coomer, SVP for Products at DDN, emphasizes the necessity for an AI storage system to support all stages of the AI workload cycle, from ingest and preparation to deep learning, checkpointing, and post-processing. Coomer references a whitepaper by OSTI. GOV titled "Characterizing Machine Learning I/O Workloads on Leadership Scale HPC Systems, " which studied the darshan logs of over 23, 000 HPC ML I/O jobs on the Summit supercomputer over the course of a year. The whitepaper notes that ML workloads generally exhibit small read and write access patterns, with an emphasis on small file reads. However, the study also observes that ML workloads generate a significant number of small file reads and writes. Coomer's blog includes a chart displaying the balance between read and write IO calls, with small calls (less than 1MB) being predominant. The authors of the whitepaper advocate for the development of storage solutions that can handle the diverse I/O patterns expected from future HPC ML I/O workloads.

Coomer compares the IO capabilities of different QLC flash systems using NFS with the DDN AI400X2 storage, highlighting the advantages of DDN's system. He contrasts a part-rack DDN AI400X2 system, capable of delivering 800GBps of write bandwidth, with a competitor's system that requires 20 racks to achieve the same performance. Coomer emphasizes that the measured write performance number for the DDN system comes from customer feedback rather than just data sheets. Various QLC flash/NFS systems are evaluated to provide additional context. For instance, a VAST Data Lightspeed storage nodes configuration offers 50GBps from 44RU, requiring approximately 16 of these configurations to achieve 800GBps write speeds. Meanwhile, a newer VAST Data Ceres system provides 680GBps from 14 racks, equating to around 17 racks for 800GBps write speeds. Another example is the Pure Storage FlashArray//C60, which offers up to 8GBps throughput from its 6RU chassis, suggesting that approximately 14 racks would be necessary for 800GBps write speed. According to Coomer, scale-out NFS system architectures are inefficient due to their complexity. DDN's Exascaler AI400X2 system simplifies the architecture by eliminating server-interconnect-buffer complexity, as the client nodes possess knowledge of the data location. Coomer argues that this system can provide the required balanced small read and write IO performance for ML workloads, such as generative AI, as supported by the OSTI. GOV research. Additionally, compared to alternative QLC flash/NFS systems, DDN's solution requires significantly less rack space, resulting in reduced power and cooling requirements for AI data centers utilizing DDN's equipment.