The vocabulary of solid-state storage, defined the way the terms are actually used in 2026. Where a definition is best understood in practice, we link the review or coverage where the concept does real work. New to SSDs entirely? Start with the SSD Reference Guide.
Interfaces and form factors
NVMe: The protocol built for flash, replacing SATA’s disk-era command model with massively parallel queues over PCIe. Every performance SSD today is NVMe.
PCIe Gen4 / Gen5 / Gen6: The bus generations that set an SSD’s speed ceiling; each doubles bandwidth. Gen4 tops out near 7.5GB/s, Gen5 near 15GB/s, and Gen6 enterprise drives are in mass production at 28GB/s and beyond.
M.2: The gumstick form factor of consumer SSDs, most commonly 2280 (22mm x 80mm), connecting over PCIe.
U.2 / U.3: 2.5-inch datacenter form factors, now yielding to EDSFF in new platforms.
EDSFF (E1.S / E3.S / E3.L): The enterprise form factor family designed for flash rather than inherited from disks. E3.S has become the default slot in AI and storage servers; see the KIOXIA CD9P-R for a current E3.S drive in the lab.
SATA: The legacy interface, capped near 550MB/s. Relevant only for old systems and some boot use.
NAND and media
NAND flash: The non-volatile memory inside every SSD, organized in pages and blocks, written in pages and erased in blocks.
SLC / MLC / TLC / QLC: One, two, three, or four bits per cell. More bits mean more capacity per dollar with lower endurance and speed. TLC is the performance mainstream; QLC owns the capacity tier, up to 245TB in one drive.
pSLC (pseudo-SLC) cache: A slice of TLC/QLC flash run in one-bit mode as a fast write buffer. Burst speeds come from this cache; sustained speeds are what remain when it fills, which is why our reviews test past it.
3D NAND / layer count: Stacking cells vertically; modern flash exceeds 200 and now 300+ layers, with 375-layer wafer-bonded NAND announced.
HBF (High Bandwidth Flash): An emerging stacked-flash class positioned between HBM and SSDs for AI memory tiers; it received its first open spec in 2026.
Performance and endurance
IOPS: Input/output operations per second, the measure of small-block random performance. Block size matters when comparing: enterprise SSD IOPS are conventionally quoted at 4K, so treat figures quoted at other block sizes carefully.
Latency: Time to complete a single operation; the metric that most affects how fast a system feels. Flash wins over disk here by orders of magnitude.
Sequential vs random: Large contiguous transfers vs scattered small ones. Marketing leads with sequential; real workloads are usually mixed.
DWPD: Drive writes per day, the enterprise endurance rating over the warranty period. Mixed-use drives like the Micron 9550 MAX carry higher DWPD than read-intensive models.
TBW: Terabytes written, the consumer endurance rating.
Wear leveling: Controller logic spreading writes evenly across flash so no block wears out early.
Garbage collection / TRIM: Background reclamation of deleted-data blocks, and the host command that tells the drive which blocks are free.
Overprovisioning: Flash reserved beyond stated capacity for endurance and performance consistency.
Architecture and features
Controller: The processor managing flash, the host interface, caching, and error correction. The single biggest differentiator between drives; see the SanDisk WD_BLACK SN8100 for what controller tuning delivers.
DRAM / DRAM-less / HMB: Onboard memory for the address map, or Host Memory Buffer borrowing system RAM in value drives.
Power-loss protection (PLP): Capacitors that flush in-flight data on power failure; standard in enterprise, absent in consumer drives.
NVMe namespaces: Logical subdivisions of one physical drive, used in enterprise deployments.
ZNS / FDP: Zoned Namespaces and Flexible Data Placement, host-guided data placement schemes that reduce write amplification in hyperscale deployments.
Emerging and adjacent
CXL: A coherent interconnect over PCIe for memory expansion, pooling, and tiering; CXL 4.0 IP now reaches 128 GT/s.
GPU-initiated storage: Letting GPUs issue reads directly to NVMe drives rather than routing through the CPU, formalized in NVIDIA’s SCADA model.
KV cache offload: Moving LLM inference context from GPU memory to flash tiers, one of the workloads reshaping enterprise SSD requirements.
HAMR: Heat-assisted magnetic recording, hard drive technology relevant here as the capacity floor SSDs compete against; covered in our Exos M 30TB review.
For current buying picks across every category, see the Leaderboard.
Amazon