Why SSD Benchmarks Look Fast but Sustained Writes Slow Down

Cet article est disponible en anglais.

Almost every storage sourcing project starts with the same number: a sequential figure from a benchmark, on a new drive, in the first seconds of the run — and it is the least predictive number in the datasheet. Copy a full card to a portable SSD and the drive leaves the conditions that produced it.

This article covers the drive side: cache, mapping memory and heat. Interface and cable behaviour belong to the link.

Where the Peak Number Comes From

Mainstream TLC NAND stores three bits per cell, so programming a page needs several verify passes across eight voltage states. Writing the same block in SLC mode — one bit per cell — is far faster.

Controllers use that difference deliberately, holding a pool of blocks in SLC mode and routing incoming writes there first. A short benchmark never leaves that pool: what it reports is the speed of the cache, not of the NAND as it ships.

Firmware later folds that data into TLC in the background. It is written twice, which costs endurance, and folding competes with whatever the host does next — one reason a drive feels sluggish for minutes after a large copy ends.

Why the Cache Is a Moving Target

Cache size is a firmware policy, recalculated from capacity, fill level and temperature, and it shrinks as the drive fills: a half-full drive absorbs less of a burst than an empty one, and a nearly full drive may write to TLC from the first block. It is also smaller on low-capacity SKUs — a 512 GB unit and a 2 TB unit will not sustain the same burst.

Neither is a defect; both are invisible in a datasheet quoting one figure.

The Three Phases of One Large Write

The trace is predictable.

Phase	Controller behaviour	What the host sees	What to record
Burst	Data lands in SLC-mode blocks; mapping updated	The rated figure, flat	Volume before the break
Transition	SLC pool exhausted; writes go to TLC while folding continues	Speed falls in steps	The drop point
Steady state	Direct TLC programming, garbage collection interleaved	A lower line, roughly flat	Post-cache speed and ambient

A fourth phase appears on used drives: as free blocks run out, garbage collection must erase before it programs, and the steady state drifts down — the difference between the second copy and the tenth.

DRAM-less Controllers and HMB

A controller with local DRAM keeps the flash translation layer — the table mapping logical blocks to physical NAND pages — in its own memory, so a lookup is a memory read rather than a flash read.

A DRAM-less controller removes that chip, holds the hottest part of the table in SRAM and, on NVMe drives, borrows a slice of system memory through the host memory buffer feature.

The window is small. Against the table of a large drive, many lookups still miss and are fetched from NAND. Sequential traffic hides this; random traffic does not.

It depends on the host. The buffer is negotiated with the operating system; where it is reduced, the controller falls back to NAND lookups — same drive, slower behaviour, no datasheet change.

The enclosure removes it. A portable SSD sits behind a USB-to-NVMe bridge; the host sees a UASP mass-storage device, not an NVMe controller, so no buffer is negotiated, and a DRAM-less drive runs on the smallest mapping window in the chain.

None of this makes DRAM-less wrong: it suits cost-sensitive products with sequential traffic.

Temperature: The Second, Later Cliff

The controller runs a physical layer, ECC and a translation layer continuously, and NAND dissipates under load. A portable drive at rated speed burns several watts in a housing a few millimetres thick; past defined junction temperatures, firmware inserts delays or drops to a lower power state.

Two details decide whether a product holds speed in the field: the thermal path — a pad to the shell, metal rather than plastic — and the environment, since a drive in a jig is not the drive on a bench.

Ambient ratings are misread here. A declared 0 °C to 70 °C range describes where a product may be used, not what it delivers there. What governs a purchase is the sustained figure at a stated ambient, housing as it ships.

Writing Sustained Write Into the Specification

Most RFQs ask for one speed. A specification that survives a real workload fixes conditions and asks for four things.

Preconditioning. Fill level, whether the drive was secure-erased, and time at the test ambient.
Workload. Block size, queue depth, one continuous transfer, and a volume several times the expected cache.
Three numbers. Burst, post-cache steady state, and volume written before the drop — the third shows whether a transfer reaches the slow part.
A curve at two ambients. Sustained speed against volume, at room temperature and at the shipping enclosure's ambient.

Then set pass or fail against the job. Where operators offload full cards between shoots, the tail dominates, so the floor belongs on steady-state speed; where a product sells on short bursts, the peak is fairer. Ask for a minimum post-cache speed per capacity SKU, and inspect every lot with one script.

How Our Own Line Is Quoted

The 2.5-inch SATA model in our internal SSD line is rated at 550 MB/s across 128 GB to 2 TB; the PCIe 4.0 M.2 model is rated at 5000 MB/s across 256 GB to 2 TB. In the portable SSD line, the PS5005 is rated at 2000 MB/s over USB 3.2 Gen 2x2 in a 35 g aluminium body, where the shell itself spreads heat.

Those are peak figures under our test conditions; sustained figures come per SKU on request, with preconditioning state, ambient and volume written before the drop.

FAQ

Is a lower peak with a flatter curve better?

Below the cache size the higher peak wins; above it, the steady-state figure and the drop point decide the total.

Does a full drive always slow down?

It slows down sooner. Less free space means a smaller SLC pool and less room for garbage collection, so writes reach TLC earlier. A half-full drive is a fair test; an empty one is not.

Can firmware remove the drop?

It can move it. Firmware sets how much space is held in SLC mode, how aggressively folding runs and where thermal thresholds sit — each trades speed against endurance.

Is a DRAM-less drive always slower?

No. It is competitive on sequential and light random traffic; the gap opens on sustained random writes and inside enclosures.

What to Take to the Next Quotation

The headline figure is a property of cache and interface; what a customer experiences is NAND throughput, mapping memory and heat. Asking for the second needs no laboratory.

Shenzhen LingC Industrial Co., Ltd. manufactures internal and portable SSDs, memory cards, card readers, USB flash drives and hubs for OEM and ODM programmes. The company was registered in 2023, and its engineering and production team brings more than ten years of storage manufacturing experience. Our quality management system is certified to IATF 16949, and we support private-label specification sheets that state sustained performance alongside peak figures for each SKU.

Cet article est également publié sur le blog lingctech.com.Voir l’article original