Program/erase cycle
One complete program and erase of a NAND erase block, a core endurance metric.
A host write is not a complete P/E cycle. An endurance rating depends on retention and error-correction conditions; it is not directly a TBW warranty. [2]
Programming can only raise a cell's threshold voltage, so a cell is reprogrammed only after an erase, and because the cells of a block share one substrate the whole block is erased. One cycle is therefore counted per block. Each cycle wears the tunnel oxide, shifting and widening the threshold-voltage distributions until they overlap. [2] [4] Survey figures are about 10,000 cycles for 5x-nm MLC, 3,000 for 1x-nm MLC and 1,000 for 1x-nm TLC. [2]
Pages of one block wear at very different speeds, so the weakest page decides when the block is unreliable. Sparing the weakest pages from hot data extended lifetime by about 10% on one 30-nm-class MLC chip and about 50% on another. [4] In a production data-centre fleet, the raw bit error rate grew smoothly with cycles past the rated limit, uncorrectable errors grew roughly linearly, and drive age mattered independently of cycles. [5]
A block of 128 pages of 8 kB holds 1 MiB, so 3,000 cycles allow about 3,000 MiB of programming per block, and the host data written is that volume divided by write amplification.
Relations
- related
- NAND flash, Wear leveling, DWPD, Planar NAND, NAND erase block, NAND dwell time, NAND self-recovery, NAND thermal annealing, MLC SSD, Dynamic wear leveling, TLC SSD
Advantages
- It gives a simple per-block wear counter that controllers compare across blocks, the basis of wear levelling. [2, 4]
Limitations
- The rated limit is a certification that is hardly sufficient to judge the real endurance of a block. [4]
- Accelerated tests showed a rapid error increase only at about three times the rated limit. [5]
Citations
- Cai et al. Error Characterization, Mitigation, and Recovery in Flash-Memory-Based Solid-State Drives. 2017, sections 2.1 and 2.3.
- Jimenez, Novo and Ienne. Wear Unleveling: Improving NAND Flash Lifetime by Balancing Page Endurance. USENIX FAST 2014.
- Schroeder, Lagisetty and Merchant. Flash Reliability in Production: The Expected and the Unexpected. USENIX FAST 2016.
Further reading
Sources & caveats
- The quoted limits and field results belong to planar-era generations and are not a rating of current devices; the survey's cycle figures are cited from earlier work.
- The field study has co-authors from Google, whose fleet supplied the data; drives report averages per drive, not per block, with average cycle counts of 185 to 949.
- The survey is university-led with one drive-maker co-author; the EPFL study is university-only.
- The lifetime gains assume up to 10% bad blocks, a 1e-4 bit-error-rate limit and 60% of writes to the hot partition.
- Arithmetic: 128 x 8 kB = 1 MiB; no source gives write amplification, so TBW cannot be derived from the rating alone (3,000 x capacity at WA 1, 1,000 x at WA 3).