An SSD's SLC Write Cache Changes Which Transfer Size Its Peak Rate Describes

SLC mode provides a different first NAND destination

An SLC-mode buffer operates NAND with one bit per cell. Samsung's 860 QVO whitepaper identifies QLC as four bits per cell. Staging writes in SLC mode provides a different first NAND destination before denser primary storage.

Samsung's 2013 TurboWrite paper describes the historical 840 EVO as simulating SLC on part of its NAND. Host writes enter that buffer first; during idle periods, data moves to the primary storage region. With consecutive writing and no idle time, its buffer eventually fills and subsequent writing goes directly to the main area.

This explains how a high rate can describe an initial interval without describing a longer transfer. Movement and reuse of the fast region depend on the implementation; the historical paper's behavior is not a universal controller policy.

References: [SC2][SC3]

Keep the buffer size and rates attached to their own product

Samsung's June 2019 860 QVO client-PC whitepaper describes variable SLC buffering. For its 1 TB variant, Table 2 lists 42 GB total: a 6 GB default region plus a 36 GB Intelligent region. That historical 860 QVO allocation does not establish 42 GB available at an arbitrary write's start.

Separately, printed page 5 of the 870 QVO datasheet, Revision 1.1, identifies Client PCs as the application and lists sequential writing up to 530 MB/s using Intelligent TurboWrite. Footnote 3 states post-TurboWrite write performance of 80 MB/s for the 1 TB variant and 160 MB/s for the 2, 4 and 8 TB variants.

Its sequential measurements use CrystalDiskMark v. 5.0.2. Footnote 2 identifies the test system and says performance can vary with firmware, hardware and configuration. These are Samsung's qualified product figures, not measurements performed for this article or a current enterprise recommendation.

Combining the 860 QVO's 42 GB region with the 870 QVO's rates would create a product prediction neither source supplies. The arithmetic below instead uses entirely invented inputs.

References: [SC2][SC1]

A wins at 10 GB; B wins at 100 GB

Consider two fictional SSD byte-service models. Use decimal GB: 1 GB = 1,000,000,000 bytes, with rates in GB/s. A starts each compared transfer with exactly 20 GB of unused fast region, serving those bytes at 1 GB/s; remaining bytes are served at 0.2 GB/s. B is stipulated to serve the entire compared transfer at a constant 0.4 GB/s. These are not real drive specifications.

Bytes are continuously available for service. Assume no overlap of fast and slow service, refill, concurrent work, protocol or host delay, or other transition cost. Each volume comparison begins independently with A's stipulated 20 GB available; the 10 GB and 100 GB cases are not consecutive arrivals.

For 10 GB, A uses only its fast region: 10 / 1 = 10 seconds. B takes 10 / 0.4 = 25 seconds. A serves that volume sooner.

For 100 GB, A's first 20 GB take 20 / 1 = 20 seconds. Its remaining 80 GB take 80 / 0.2 = 400 seconds, for 420 seconds in total. B takes 100 / 0.4 = 250 seconds. B serves that volume sooner despite its lower peak.

Dividing all 100 GB by A's 1 GB/s would claim 100 seconds by assigning fast service to bytes beyond the initial 20 GB allowance. The two-part calculation exposes that mistake. These are calculated byte-service intervals, not measured application completion times or durability evidence.

The ranking changes at a calculable volume

Let V be the transfer volume in decimal GB. Under the same assumptions, A's service time in seconds is min(V, 20) / 1 + max(V - 20, 0) / 0.2. B's is V / 0.4.

Above 20 GB, A's expression becomes 20 + 5(V - 20) = 5V - 80; B's becomes 2.5V. They are equal at 32 GB, where both take 80 seconds. A is faster for smaller positive volumes in this model; B is faster above 32 GB. That crossover belongs only to these invented inputs.

If the requirement were 100 GB within 300 seconds of byte service, B would pass that limited screen and A would not. An actual application deadline includes any work omitted from the model, so passing this calculation does not establish application readiness.

The next arrival does not inherit the first cache state

Samsung's consumer-SSD FAQ describes Intelligent TurboWrite moving data from the SLC buffer to multi-bit cells when the SSD is neither reading nor writing, freeing buffer room. That description supplies no universal movement rate or recovery time.

For recurring staging writes, the relevant quantity is room available when each occurrence starts. Earlier writes can still occupy the region; intervening reads or writes can change background movement. A's stipulated 20 GB cannot automatically be reused for every occurrence.

Equal daily byte totals can arrive as separated small stages or as one concentrated run. Those timelines offer different opportunities for buffer reuse. Background relocation and other activity can also invalidate the model's separate, constant rates. A published maximum region size cannot establish its initial available capacity.

References: [SC4]

Separate SLC buffering from DRAM, reserve capacity and reclamation

Samsung's 2014 power-loss-protection note describes capacitor energy supporting a transfer from volatile DRAM to NAND after external power disappears. An SLC-mode NAND buffer is already nonvolatile media. Its size and speed do not establish protection for pending volatile writes or the metadata needed to locate data. The related power-loss-protection guide owns that assessment.

Manufacturer-set over-provisioning describes capacity withheld from host access for controller use, as Kingston's March 2019 explanation states. SLC buffering describes NAND's use as a fast staging region. These functions differ; neither label establishes that the physical regions are equal or disjoint. An OP percentage cannot supply a missing SLC-buffer allowance.

Garbage collection concerns preserving valid pages and reclaiming erase blocks. The linked garbage-collection article explains that work and fresh versus workload-conditioned performance. Finishing an SLC region introduces another question: what placement path and NAND mode serve subsequent bytes? A post-cache rate is not automatically a whole-workload steady-state result.

References: [PL1][OP1]

Compare evidence for the whole recurring write stage

Define the bytes that must reach the selected SSD during the write window and when that window recurs. Match the offer's evidence to that volume, arrival pattern and completion boundary. Retain the exact model, capacity variant and applicable firmware and test-system qualifications with any manufacturer rate.

Ask the storage owner what establishes the initially available fast region and performance across the required volume after that region is used. Evidence from one short transfer leaves a longer or recurring stage unresolved. That omission does not prove a defective drive; it identifies the missing basis for the buying decision.

The useful choice can favor A for one volume and B for another without declaring either design superior for every use. The related performance guide supplies the IOPS, throughput and latency distinctions. Endurance ratings and installed capacity remain separate requirements; neither supplies the missing duration evidence.

Carry selected equipment into the physical installation

Once the storage and platform owners have confirmed the hardware selection, Leviathan's rack-and-stack service can address agreed equipment mounting, IT cabling, labeling and physical installation records. The service and its project inquiry are linked below.

Share a nonconfidential equipment summary, released rack and connection plans, site constraints and required physical handoff. Keep the storage-selection owner and equipment supplier identified. SSD performance evaluation, firmware changes and software administration require separately appointed scope; an installation record does not establish the write-stage result.

References: [L1]