Research Note: NetApp introduces new AFF A-Series hardware targeted at Artificial Intelligence

NetApp has introduced new AFF A-Series storage solutions that are being targeted at the requirement for high performance and low latency in AI training and inference use-cases. Is this an incremental evolution of the hardware or something fundamentally different?

Background

The AFF (all-flash FAS) series of ONTAP-based storage solutions have been around for almost ten years. The first AFF systems (the AFF8000 series) were introduced in 2015 and have been expanded and updated ever since. The concept of a dedicated all-flash system is quite simple. If the storage operating system knows the only devices deployed are flash based, then the media management algorithms can be optimised accordingly. There is no need, for example, to cater for the relatively long latency times of spinning media.

Product Diversity

We’ve covered the expansion of the ONTAP family in a series of posts over the last five years. In November 2019, NetApp introduced the All-SAN Array (ASA), to complement the traditional mixed protocol platform of AFF. The ASA was relaunched in May 2023.

In February 2023, NetApp announced the C-Series, a capacity-optimised solution using QLC NAND storage. An ASA C-Series was announced in November 2023, completing the line-up of standard (A-Series) and cost optimised (C-Series) AFF and ASA solutions.

AFF Updates

The new AFF A-Series announcements introduce three new products, with incremental capacity and performance capabilities. The AFF A70, AFF A90 and AFF A1K all provide up to 15.5PB of effective capacity. The A70 and A90 scale from 68TB to 3.7PG raw, while the A1K scales from 91TB to 3.7PB raw. NetApp uses an effective capacity ratio of 5:1 when translating from raw capacity values.

The design of the A70 and A90 solutions differ slightly from the A1K in that they are deployed as a 4U base unit with 48 drives. The A1K by comparison is a 2U+2U dual controller solution with separate 2U drive shelves. This design choice was made to enable higher performance on the A1K, which has greater airflow/cooling requirements.

Performance

NetApp claims 2x performance improvements for the new hardware over previous AFF solutions, with up to 40 million IOPS and 1TB throughput. Probably more interesting for customers is the translation of that performance into financial savings. NetApp claims the new systems will offer 50% lower $/IOPS, 55% lower $/throughput, up to 35% $/TB effective and 45% lower $/density.

These financial claims are based on a new storage efficiency ratio of 3.9:1 compared to the previous 2.27:1 and represent technical improvements in data reduction techniques.

Architecture

As with all modern storage arrays, performance improvements are driven by the adoption of new processor architectures. The latest A-Series solutions use Intel Sapphire Rapids 4^th Generation Xeon processors (a move that Pure Storage made last year with R4), which includes Intel Quick Assist Technology to offload some of the compression and deduplication calculations.

The choice of vendor and chipset is interesting, with clear differences, for example, between Intel and AMD processor families, especially with respect to I/O. The negative side to dependency on the processor vendor is the cadence of product releases, which are directly tied to the arrival of new processor designs.

The Architect’s View®

Every hardware vendor is looking for an AI angle that makes their solutions more relevant in the current market. NetApp is no exception, with a new set of reference architectures called AIPod, which uses NVIDIA DGX servers.

Is the new storage hardware “AI-specific”? From our view, not really. However, we would suggest that the latest iteration of AFF solutions is certainly “AI-friendly”, and that’s a good thing. Performance is improved, while the cost of ownership is reduced on the basis of metrics that are relevant to AI – capacity and performance.

NetApp would like us to think about their solutions with reference to the concept of “Intelligent Data Infrastructure”. Certainly, these solutions already address many of the aspects NetApp considers under the IDI moniker, including data and protocol ubiquity (any data, anywhere), active data management and adaptive operations.

If AI is the future of computing, then a solid storage foundation (especially on-premises) will continue to be essential. It’s quite interesting to reflect on the staying power of consolidated and centralised storage personified by SAN technology, that still has a future in the enterprise. From NetApp’s perspective, that is good news, as storage hardware still drives the overwhelming majority of revenue for the company.