DDN doubles performance of high-end AI array

DDN has boosted the sequential bandwidth performance of its high-end A3I AI400X all-flash array almost twice over with an AI400X2 upgraded system in the same chassis.

The A3I arrays are based on DDN’s EXAScaler arrays running Lustre parallel file system software. They are optimised for artificial intelligence (AI) applications and support Nvidia’s DGX servers — the A100, the Pod and the SuperPod. 

Update – AI400X2 performance data added. November 11.

Dr James Coomer, SVP of products at DDN, provided an announcement quote: “We designed this next generation of A3I solutions to give customers the most efficient, scalable and reliable AI data platform.” Well, what else would it be for?

DDN AI400X2 array.

As well as the faster performance, the AI400X2 features so-called EXAScaler Hot Node integration, a streamlined user interface and enhanced security. The Hot Node technology, co-developed with Nvidia, caches files on the accessing client’s local NVMe flash storage to speed data access.   

The better UI capabilities include granular visibility into AI workloads and client GPU-level utilisation to simplify configuration management and system monitoring, and optimise end-to-end performance. On the security front there are client-side encryption and role-based access control configurations and policies.

Doubled performance

The new AI400X2 uses the same enclosure as the existing AI400X and DDN’s announcement claims it has ”twice the performance of the previous generation appliance in the same footprint”. This refers to its sequential bandwidth as the table below, comparing it to the existing A3I models, illustrates;

Blocks and Files table.

The added performance comes from a change to an internal PCIe Gen 4 switch fabric with up to 288 lanes of PCle Gen 4, plus up to 64 lanes providing client connectivity. We understand that DDN actually measured 94GB/sec sequential read and 70GB/sec sequential write bandwidth, but put more conservative numbers in the datasheet.

The new system also uses four of the latest gen 3 Xeon CPUs and has a larger memory footprint than the AI400X.

EXAScaler base systems

We decided to check the EXAScaler product family to try and discover a base system for the AI400X R2. There are four EXAScaler EXA6 products: ES200NVX, ES400NVX, ES7990X and ES18K.

DDN EXAScaler information.

We can see from our own table above that the AI7990X is clearly based on the ES79909X, the AI200X on the ES200NVX and the AI400X uses the ES400NVX as its base. However there is no EXAScaler base system for the new AI400X2 array. This leads us to expect that an EXAScaler ES400NVX refresh is coming. And, we understand, indeed one is.