NAND 層數與每儲存格位元數：NAND 密度的極限在哪裏

Two numbers define every solid-state drive you buy: how many layers the manufacturer can stack, and how many bits each cell can hold. Both have been rising for a decade. Both are now approaching a wall, and the walls are not the same wall.

Understanding which one is binding explains why the price of flash has stopped falling as fast as the marketing implies, and why the drive you buy for a write-heavy workload is not the drive with the best headline density.

Part One: Two levers, different physics

Adding layers is a manufacturing problem. Each generation etches more floors into the same wafer, which spreads the lithography and deposition cost across more cells. The economics are close to linear in layer count, and the engineering difficulty is in etch depth and uniformity — deep, high-aspect-ratio holes that must be straight to within a few nanometres.

Adding bits per cell is an analogue problem. A cell that stores four levels must distinguish them reliably against noise, retention drift, and temperature. Five levels is harder, and the margin between levels shrinks faster than the error-correction budget can absorb.

The first lever is nearly free per bit. The second is increasingly expensive, because every extra level demands more precise programming, slower writes, and more over-provisioning to cover the failure rate.

Layer count is a scaling lever; bits per cell is a precision lever, and precision has a floor that stacking does not.

Part Two: The endurance bill

Every bit of density is paid for in write endurance. A single-level cell tolerates around a hundred thousand program-erase cycles. Triple-level, the volume workhorse, manages about a thousand. Quad-level drops to a few hundred. Penta-level, the current frontier, lands near a hundred.

This is not a minor specification. It decides what the drive is for. A QLC drive is excellent for read-mostly data — media libraries, model weights, archives — and unsuitable for a write-ahead log, because the log will exhaust the cells in months.

The industry's answer is over-provisioning and caching: keep a fast SLC region as a write buffer, fold it into QLC in the background. That works, and it is why a QLC drive can absorb bursts. It does not change the steady-state endurance, only the shape of the curve.

There is a second-order cost that rarely appears in the datasheet: write amplification. As cells get denser, the drive must read-modify-write larger blocks, so a small random write consumes more cell cycles than its size suggests. At five bits per cell, the amplification can be severe enough that the rated endurance is optimistic for mixed workloads.

Part Three: Why the cost curve flattens

For a decade, cost per terabyte fell by roughly thirty percent a year. That has slowed, and the reason is that both levers are now in their expensive region.

Layers continue to rise, but each generation needs more deposition steps, more etch time, and more capital equipment. The cost per wafer rises almost as fast as the bits per wafer, so the cost per bit falls slowly. This is why fabs talk about layer count with pride and about cost per bit with caution.

Bits per cell has effectively stalled at four for mainstream volume, with five shipping in limited products. The endurance penalty and the controller complexity mean that the cheapest drive per terabyte is not the densest one — it is the one whose density matches the workload's write rate.

The result is a market that segments by workload rather than by capacity. The same silicon generation produces a cheap archive drive and an expensive mixed-use drive, and the price difference is endurance, not bits.

Part Four: What this means for system design

Three practical consequences.

Stop assuming storage gets cheaper. Capacity planning that extrapolates a thirty percent annual decline will over-provision in year three and under-budget in year five. Model a slow decline and revisit.

Separate the write path from the read path. Put logs and metadata on high-endurance media and bulk data on dense media. The blended cost is lower than either uniform choice, and the failure modes are clearer.

Treat endurance as a first-class capacity. A drive has two capacities: bytes and total bytes written over its life. The second one is the one that runs out first in write-heavy systems, and it is the one most capacity plans ignore.

What is actually transferable

The general pattern is that density and endurance trade against each other in every storage medium, and the trade gets worse as density rises. Magnetic disks went through the same curve: areal density gains eventually required shingled writes and host-managed zones, pushing complexity onto the software.

The lesson for anyone building a data platform is to stop treating storage as a homogeneous pool. It is a set of tiers with different endurance, different write amplification, and different failure modes. The systems that age well are the ones that assign data to tiers by access pattern rather than by size, and that budget for the day the cheap tier stops getting cheaper.

SOL: HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Creating this takes real effort — thank you for the support.