Why Raw NAND Flash with Hardware-based ECC is the Way to Go
5 Reasons for Ditching Managed NAND
Everyone knows that NAND has challenges: from factory bad blocks and spontaneous bit failures to endurance limits, etc. That’s why a few years ago managed NAND (NAND flash plus an integrated controller) seemed to be the answer, offering the density of raw NAND, while mitigating many of its inherent limitations. What many device manufacturers may not realize is that the management hardware comes with significant costs, both in terms of dollars per part as well as design limitations. In the world of tradeoffs in which every system designer lives, there are solid technical reasons to consider using raw NAND and leaving the management to software instead. While there are clear commercial advantages for Datalight (as a provider of vendor-neutral software-based flash management) to advocate this approach, we also believe that there are strong technical reasons that flash silicon vendors would do well to embed ECC capabilities into their NAND flash devices rather than relying on ‘total hardware’ solutions such as eMMC or other complex and costly controllers. Beyond the benefits outline below, this approach would allow the flash manufacturers the ability to continue to differentiate their products from others in the industry.
The cost of managed NAND parts is coming down, but the stuff still sells at a premium over its raw NAND brethren.
2. Flash Optimization
There are many new features of NAND available to us today. Performance features such as cached reads, multi-plane operations, concurrency, and others are becoming invaluable to keep performance at the ever-increasing demands of portable media. The Open NAND Flash Interface (ONFI) has defined a standard method to query the capabilities and characteristics of NAND flash which can be put to use by both software and hardware systems. A software media manager offers the flexibility to take advantage of the most current flash memory features and put them to use efficiently, or to avoid certain features that may be unproven or problematic. A software solution will allow a developer to take full advantage of the media’s characteristics and features unburdened by the indirection or inability for the hardware to expose them.
Software (in general) is easily inspected and validated. Features such as wear-leveling move data around the flash device to optimize its life expectancy. Without the ability to inspect source code, a managed NAND solution makes it difficult to validate wear-leveling operation and/or characterize its effect on performance and reliability. Hardware implementations are often generalized to suit a majority of use cases, while a software solution is easily tailored to the specific use case during development.
Speaking of use cases, there are many system features that are not available to hardware that may make a generalized hardware solution less advantageous to a specific use case. For example, system idle time can be used to improve the media performance by scheduling background cache operations and compaction to occur then. Coordination between the file system and flash media manager can further optimize operations by freeing space when it will no longer be needed and having the media manager code cache certain regions of the flash where meta data might be held. Migrating flash management features to hardware removes this ability to coordinate with other components of the software stack, such as file systems.
Lest you think we believe that everything is better left to software, consider error detection and correction (EDC). Error rates are increasing substantially as flash manufacturers push the limits of physics. Errors can be introduced externally by heat or other radiation, during writes or reads of data, and even to data that was successfully written at a different time. Historically SLC NAND flash required only a single bit error detection and correction (a hamming code is usually sufficient), while MLC parts require minimally four bit EDC. As the die sizes continue to shrink, error rates will continue to increase, even for SLC flash.
Calculating the codes to detect and correct such errors is getting increasingly complex and solving such a solution in software for higher-bit EDCs (above 4-bit) is time consuming and often unacceptably slow. Hardware ECC is a necessary requirement for systems with high EDC requirements and where performance is a concern.
Many of the processors on the market today are incorporating EDC in their NAND controllers. Choosing one of these processors (e.g. TI OMAP 35xx) in combination with raw NAND flash and software management can give you the high-performance EDC to handle next generation flash while maintaining the design flexibility that a software manager provides.
Flash manufacturers have much to gain by adding ECC code into their NAND flash parts. They know better than anyone what kind of ECC is necessary for a specific part and by adding just that one piece of hardware to their offerings, rather than the jack of all trades, master of none approach of complete flash management, they will better serve the markets.
In short, features should reside where they can be handled most efficiently; ECCs belong in hardware, other flash management functions belong in software. While managed NAND certainly has its place and its appeal in the market, we believe the best combination of value, performance and flexibility lies in using a combination of raw NAND and hardware with built-in ECC capabilities.