A Computerworld article last week reported on a reseacher’s prediction that shrinking the size of NAND flash memory for solid-state drives (SSDs) may cause the technology to lose significance altogether.
In a lecture delivered at this week’s Usenix Conference on File and Storage Technologies, Laura Grupp, a graduate student at the University of California, San Diego, argued that as flash memory is manufactured at smaller geometries, data errors and latency would increase. The idea behind shrinking the transistors is to boost capacity, which translates into lower cost per gigabyte. But that pushes performance and reliability in the opposite direction.
Grupp wrote about the phenomenon in a study titled The Bleak Future of NAND Flash Memory. “While the growing capacity of SSDs and high IOP rates will make them attractive for many applications, the reduction in performance that is necessary to increase capacity while keeping costs in check may make it difficult for SSDs to scale as a viable technology for some applications,” she wrote.
Grupp, along with John Davis of Microsoft Research and Steven Swanson of UCSD’s Non-Volatile Systems Lab, tested 45 types of NAND flash chips, spread across six vendors and multiple transistor geometries (between 25nm and 72nm). The researchers found that write speed for flash blocks had high variations in latency. In addition, they also discovered wide variations in error rates as the NAND flash wore out. Multi-level cell (MLC), and especially triple-level cell (TLC) NAND created the worst results, while single-level cell (SLC) performed the best.
Grupp, Swanson and Davis extrapolated the results to 6.5nm technology, which is the size NAND transistors are expected to be in 2024. At that size, the researchers estimate that read/write latency will double in multi-level flash, with triple-level suffering 2.5 times as much latency. Bit error rates are expected to increase as well, more than tripling those of current levels.
But since flash memory is a solid-state technology (versus the mechanical technology used in hard disks), SSDs will always have a natural advantage in speed and throughput. In general, reading and writing to an SSD is about 100 times faster than that of a hard drive.
Grupp concedes that even with 2024-level transistor sizes, SSDs will outperform their hard disk competition by a wide margin, 32,000 IOPS to 200 IOPS, respectively. But because of the latency and error rate issues, she believes that 6.5nm will be end of the line for flash memory.
For flash memory, there seems to be a choice of performance or capacity, but not both. This could have lasting impacts on data-intensive applications that lean heavily on IOPS performance. Without a replacement for NAND memory, performance could stall or even decline until another solid-state technology takes its place.