DRAM-less SSD

On this page 8

A DRAM-less SSD is a solid-state drive whose controller has no dedicated DRAM chip for caching the logical-to-physical mapping table, the index that tells the drive where each block of data physically sits in NAND flash. In hardware storage, this design lowers cost, power draw and board size, but it shifts the burden of address translation onto small on-controller SRAM, the flash itself or, in NVMe drives, a slice of the computer’s own memory.

The distinction matters because the mapping table is consulted on every read and write. Where the table lives determines how quickly the drive handles small random accesses, how it behaves when nearly full, and how gracefully it ages, so two drives with identical sequential speed ratings can feel very different in daily use.

How a DRAM-less SSD tracks its data

Every SSD runs a flash translation layer (FTL) that maps logical block addresses from the operating system to physical pages in NAND. A conventional design keeps this map in DRAM at a ratio of roughly 1 GB of DRAM per 1 TB of capacity, because a page-level map needs about 4 bytes for every 4 KiB of storage.

A DRAM-less controller cannot hold the full map in fast memory. Instead it keeps a small cached portion of the table in its internal SRAM, usually a few megabytes, and stores the complete table in NAND flash. When a request targets an address whose map entry is not cached, the controller must first read the mapping entry from flash, adding an extra flash read before the actual data access.

The following table shows where the mapping table can reside and what each option implies for latency.

LocationTypical size availableAccess latencyUsed by
Dedicated DRAM512 MB to 4 GBAround 100 nsMainstream and premium SSDs
Controller SRAMA few MBUnder 10 nsAll DRAM-less SSDs
Host memory bufferTypically 16 to 64 MBAbout 1 to 2 µs over PCIeDRAM-less NVMe SSDs
NAND flashFull tableRoughly 50 to 100 µsFallback for uncached entries
Locations of the mapping table in SSD designs

Host memory buffer and its role

The Host Memory Buffer (HMB) is an optional feature of the NVMe specification that lets a DRAM-less drive borrow a region of system RAM over PCIe. The operating system allocates the buffer at boot, and the controller uses it as a secondary map cache, which is much faster than fetching entries from NAND even though it is slower than local DRAM.

A buffer of 64 MB can cache map entries covering about 64 GB of data at page granularity, which comfortably covers the active working set of a typical desktop. Because of HMB, a modern DRAM-less NVMe drive often performs within a modest margin of DRAM-equipped drives in light client workloads, while DRAM-less SATA drives, which have no HMB mechanism, show a much larger penalty.

HMB depends on operating system and driver support. Current Windows, Linux and macOS releases support it, but some hypervisors, older systems and external USB enclosures do not pass it through, leaving the drive to rely on SRAM alone.

Performance characteristics and trade-offs

The weaknesses of a DRAM-less SSD appear mainly under random access across a wide address span, in sustained writes and when the drive is nearly full. Sequential transfer rates are often close to DRAM-equipped models, because sequential access reuses the same cached map entries.

The table below compares typical behavior across workload types.

WorkloadDRAM-less SATADRAM-less NVMe with HMBMain reason
Sequential readsSimilarSimilarMap entries are reused
Small random reads, narrow spanSlightly slowerNear parityWorking set fits in cache
Small random reads, wide spanMuch slowerNoticeably slowerCache misses force flash lookups
Sustained heavy writesMuch slowerSlowerMap updates compete with data writes
Nearly full driveDegrades sharplyDegrades moderatelyGarbage collection and map churn
Typical behavior compared with a DRAM-equipped SSD

Write behavior is also shaped by the SLC cache, which absorbs bursts before data is folded into TLC or QLC cells. DRAM-less drives frequently pair a modest SLC cache with QLC NAND, so once that cache fills, write speed can fall to a few hundred megabytes per second or lower.

Endurance, wear leveling and reliability

Without DRAM, the controller must flush map updates to NAND more often, because it has little space to batch them. These extra metadata writes raise write amplification, the ratio between data physically written to flash and data the host requested, which in turn consumes rated endurance (TBW) faster under write-heavy use.

Modern firmware mitigates this with coarser mapping granularity, journaling and compression of map pages, so for ordinary client workloads the endurance difference is usually small. Wear leveling and TRIM still operate normally. One reliability benefit is that a drive with less volatile state has less data at risk during sudden power loss, although power-loss protection itself depends on the specific design.

Choosing between DRAM-less and DRAM-equipped drives

The right choice depends on workload intensity and interface, not on the label alone. A DRAM-less NVMe drive is a sound choice for many consumer roles, while servers and heavy multitasking benefit from a dedicated cache.

Use caseRecommended typeReason
Office laptop or web browsingDRAM-less NVMe with HMBLight workload fits in HMB cache
Game library storageDRAM-less NVMeMostly sequential reads of large files
Operating system drive for heavy multitaskingDRAM-equipped NVMeWide random access across the drive
Video editing scratch diskDRAM-equipped NVMeSustained large writes
Database or virtualization hostDRAM-equipped enterprise SSDConsistent latency under load
Upgrading an older SATA systemDRAM-equipped SATANo HMB fallback on SATA
Decision guide by use case

Spec sheets rarely say DRAM-less explicitly. Buyers usually identify such drives by the controller model, by board photos showing no DRAM package, or by the presence of an HMB entry in NVMe identify data, which tools such as nvme id-ctrl report as hmpre and hmmin values.

Common misconceptions

  • DRAM-less does not mean unreliable. Failure rates depend far more on firmware quality and NAND grade than on the presence of DRAM.
  • The SLC cache is not a DRAM substitute. It buffers user data writes, while DRAM caches the mapping table; the two solve different problems.
  • Peak sequential speed hides the difference. Box figures in megabytes per second say little about random performance across a wide span.
  • HMB is not automatic everywhere. In a USB enclosure or an unsupported virtual machine, an NVMe DRAM-less drive loses its main mitigation.

In everyday language

An SSD needs an index to know where each piece of your data is stored, much like a library needs a catalog to find each book. A drive with DRAM keeps the whole catalog on a fast desk right next to the librarian. A DRAM-less SSD keeps only a few catalog pages on the desk and stores the rest back in the shelves.

For everyday tasks, the librarian rarely has to fetch the full catalog, so the drive feels quick. When you ask for many scattered items at once, or the shelves are nearly full, the extra trips make it noticeably slower. The trade is simple: lower price and power use in exchange for weaker performance under heavy load.

Seen in the real world

A small accounting office replaces the aging hard drives in eight office laptops with 1 TB DRAM-less NVMe SSDs that support HMB. The laptops run spreadsheets, email and a browser, and the IT technician confirms that the operating system allocates a 64 MB host memory buffer on each machine. Boot times drop from over a minute to about 15 seconds, and staff notice no difference compared with a DRAM-equipped test unit.

The same office briefly tries one of these drives in a workstation that runs a local accounting database and two virtual machines. Under that load, with the drive about 85 percent full, query times grow erratic and large file copies slow sharply once the SLC cache fills. The technician moves the workstation to a DRAM-equipped drive, keeping the cheaper models where they fit: light client workloads with modest write volumes.