NAND die stacking
NAND die stacking places multiple physical NAND dies in one package to increase capacity per package.
Package-level stacking combines separate dies; 3D NAND instead arranges memory cells in layers within a die. A package can use both approaches. [2]
The 2008 design study describes two dies sharing an I/O bus, with separate chip-enable and ready/busy signals. That is a concrete organization, not a definition of all stacked packages. More dies can expose overlapping array work without multiplying the bandwidth of shared transfer wires. [3]
Keep physical construction separate from interface geometry. ONFI groups logical units into targets, while Linux represents targets, LUNs and planes separately. Package capacity alone cannot reveal those counts or the supported command schedule. [4] [5]
Historical scope matters. The 2017 survey describes 4-16 chips per SSD, up to 16 dies per chip and 1-4 planes per die in its architectural account; these are not present-day limits. Its superpage parity example assigns one die's contribution to parity and c x d - 1 to data for c chips with d dies each. Recovery depends on that parity organization, not stacking alone. [2]
Relations
- related
- 3D NAND, V-NAND, NAND channel, Planar NAND, NAND flash, NAND target, NAND logical unit, NAND plane
Advantages
- The survey's superpage-parity design can recover one unavailable die contribution when the remaining contributions and parity are usable; stacking alone supplies no such guarantee. [2]
Limitations
- Adding dies does not by itself remove contention on a shared NAND channel. [3]
Citations
- Cai, Ghose, Haratsch, Luo and Mutlu. Errors in Flash-Memory-Based Solid-State Drives: Analysis, Mitigation, and Recovery. arXiv 1711.11427.
- Agrawal et al., Design Tradeoffs for SSD Performance, USENIX ATC 2008
- ONFI 4.0, sections 1.4.1 and 3: NAND target, logical unit and memory organization
- Linux v6.12, include/linux/mtd/nand.h: nand_memory_organization, nand_pos and nanddev_target_size
Further reading
Sources & caveats
- The 2008 paper and 2017 survey describe historical architectures, not present-day stack-height, package-count or timing limits.
- Physical packaging and controller-visible addressing are different descriptions; confirm their mapping in the selected component's data sheet.
- The error survey is university-led with one drive-maker co-author; it does not establish current packaging limits, price or yield.