It has been nearly four years since the initial announcement from Intel regarding the development of a new storage technology called 3D Xpoint. This new high-performance media was developed under a partnership between Intel and Micron, named IM Flash. 3D XPoint was a newly developed Nonvolatile Memory (NVM) based on an innovative vertical architecture featuring performance and endurance qualities that were unprecedented. In 2016, IM Flash began to produce PCIe based drives and U.2 drives constructed with the 3D Xpoint architecture. The performance metrics were impressive when compared to competing products constructed of Triple Level Cell Nand (TLC). However, the most anticipated product from the 3D XPoint family was still on the horizon, Persistent memory (PMEM), code named Apache Pass.
PMEM has the potential to alter the way we architect our applications and storage devices. The promise of 3D Xpoint based PMEM can unlock byte addressable, persistent memory at improved speeds and reduced latencies not seen from more traditional storage media. Imagine a high-performance storage media that didn’t pay a performance penalty for writes. All other types of SSD’s whether 12Gb SAS, 6Gb SATA or NVMe suffer performance write degradation under heavy write or large block IO processing. Unlike the PCIe based or U.2 3D Xpoint storage devices that were compatible with existing server hardware, 3D Xpoint (Apache Pass) PMEM was dependent on the development of a new processor technology.
In early April Intel announced the availability of their new processor family, Cascade Lake. This new generation of Xeon processors offered several key features such as Intel’s Deep Learning Boost, Advanced Performance architecture and support for Apache Pass (PMEM). It is important to understand that Apache Pass is not designed to displace current DRAM completely. Apache Pass holds great promise as a non-volatile storage media that is positioned on the memory bus of a server. The benefit of non-volatile storage located on the memory bus is its ability to offer impressive levels of storage performance and dramatically reduced latencies.
Initial capacities for Apache Pass have been announced. They range in size from 128GB to 256GB and 512GB. These devices are pin compatible with DDR4. The Apache Pass modules can be deployed in block mode or memory mode. Intel provides a deployment methodology based on a ratio of 4GB to 1GB (PMEM to DRAM). Both application and storage vendors have spent vast resources adapting their technologies for Apache Pass. Hyperconverged Infrastructure looks like a great candidate to take advantage of this emerging technology by using PMEM as a cache mechanism, tier 0 or for residence of metadata. New companies such as Vast Data are developing tiered storage products that implement both PMEM on the front side and quad level cell nand (QLC) for long tail storage. From an application perspective “in memory” processing of structured data is also a great candidate for this new technology.
Finally, there are a few caveats that should be considered in relation to Apache Pass. First, IM Flash will be dissolving their partnership soon. Both Intel and Micron will take the Apache Pass intellectual property and begin going down divergent development paths. Currently, Intel seems to have a more defined road map for Apache Pass than Micron. So the vision of the next generation of Apache Pass is murky at best. Second, production facilities for Apache Pass is very limited. This product is being fabricated at a plant in Lehi, Utah. Limited production capabilities will keep prices for Apache Pass inflated during the early adoption period. Third, competition to source Apache Pass will be intense so check with your Sanity Solutions team for an overview of availability and how these developments may affect your short and long term storage strategy.