Sources

The Throttle Lab is an illustrative model, not vendor data. The trends are real. The exact heat and power numbers belong to no real drive: every thermal resistance, heat capacity, time constant and power figure is a model value, sized so the shapes come out right, and the page labels them that way. This file lists the claims the lab makes, says where each one comes from and how strong the evidence is, and lists what was checked but not shipped.

Evidence grades: [V] vendor primary source (white paper, datasheet, vendor blog). [R] independent review or measurement. [S] standards body (NVM Express, OCP, Linux kernel documentation). [P] patent: describes a method, does not prove any shipping drive uses it. [paper] peer-reviewed or preprint paper. [F] forum post or user-posted dump, one observer. [M] vendor or third-party marketing claim. [MODEL] a lab default, not a published figure.

Checked on 2026-10-04. Primary texts re-read the same day: Samsung SSD 950 PRO white paper (section 3.3.1, Table 5), NVM Express Base Specification 2.1 (2024-08-05), the nvme-cli master and v2.11 manual pages, the Linux drivers/nvme/host/hwmon.c source, and the Microsoft Learn Get-StorageReliabilityCounter page. The research reports behind the mechanism table are in docs/research/.

The mechanisms the lab models

Setting	Claim	Grade	Source
Stepped levels	The Samsung 950 PRO uses 10 levels. Level 0 is full speed. Entry is 75°C in Table 5 (the text next to it says 74°C). Levels 1 to 7 are 87.5%, 75%, 62.5%, 50%, 37.5%, 25% and 12.5% of full performance at 75, 75.5, 76, 76.5, 77, 77.5 and 78°C. Level 8 (78.5°C) is labeled “DTT performance” with no figure. Level 9 (79°C) is labeled “Meltdown (10 MB/s)”.	[V]	Samsung SSD 950 PRO white paper, section 3.3.1 and Table 5. The lab uses 75°C from the table and notes the 74°C in the text. Level 8 has no stated percentage, so the lab treats it as a model value between 12.5% and the level 9 floor.
Stepped levels	The stepping algorithm lowers one level, measures again, and goes one more level if the temperature is not lower, until the temperature settles. The drive returns to full performance quickly once under the threshold.	[V]	Same white paper, text under Table 5 and the Figure 11 caption.
Stepped levels	Transcend lists 3 protection levels at 85, 95 and 100°C. ATP lists up to 18 steps from 85°C and says conventional designs use two or three stages, which produces abrupt drops and a drive that swings between full speed and a deep cut.	[V]	Transcend, Thermal Throttling (embedded); ATP, How does thermal throttling work in SSDs
I/O pacing	A Phison-partner SSD brand says a throttling drive typically delays I/O.	[V]	SSD brand support guide, SSD throttling
I/O pacing	Innodisk’s patented scheme checks temperature every second, computes the difference from a target, and raises or lowers the number of NAND commands. Its five stages are full speed, performance decrease, limit I/O, performance increase, full speed. The same paper says most PCIe SSDs use a simple scheme that lowers the flash clock at 85°C.	[V]	Innodisk, A Modern Approach to Thermal Throttling in PCIe SSDs (white paper)
I/O pacing	A proportional or PID response can hold a temperature near its limit with less lost performance than on/off toggling (CPU research: threshold within 0.2°C of the maximum, 65% less performance loss than toggling).	[paper]	Skadron, Abdelzaher and Stan, HPCA 2002, abstract. CPU work, not SSD work. No peer-reviewed SSD throttle-controller design was found.
Link state	Phison calls it Link-State Thermal Throttling: it reduces the PCIe link “at the PHY” internally, moving from Gen5 to Gen4, Gen3 and so on, and the drive regains a Gen5 link when the thermal event ends. First shipped as firmware 22.1 for the Corsair MP700 (Phison E26).	[V]	Phison blog (2023-07-25); TweakTown article carrying the Phison statement
Link state	The threshold on firmware 22.1 is 85°C, a bare MP700 then reads above 10 GB/s and writes about 2 GB/s.	[R]	ComputerBase, relayed by Tom’s Hardware. The lab says “mid-80s” because Tom’s is the secondary source.
Link state	Tom’s describes the Crucial T705 as throttling with “a controlled decline in PCIe link speed”.	[R]	Tom’s Hardware, Crucial T705 review
Link state	A SanDisk patent describes duty-cycling the PCIe rate between Gen5 and Gen4 on a thermal or power threshold.	[P]	US12360686B2
Link state	Whether a host tool such as lspci shows the lower rate during an event is unconfirmed. The lab shows “internal link rate” as a model readout and never says a host tool shows it.		Vendor research report, section 1.1
Link state	In the reviews we saved, only the E26 is documented with this method. Reviews of the E28 and E31T describe gradual throttling with no link-state mention.	[R]	TechPowerUp E28 and E31T engineering sample reviews
NVMe power states and HCTM	The host sets two temperatures. Above TMT1 the controller moves toward lower active power states or other vendor actions while minimizing the performance hit (light throttling). Above TMT2 it does so regardless of the performance impact (heavy throttling). Each is in kelvins, 0 disables it, TMT1 must be below TMT2, and the controller advertises a minimum and maximum. In the Feature 10h value, TMT1 is bits 31:16 and TMT2 is bits 15:0.	[S]	NVMe Base Specification 2.1, 5.1.25.1.9 (Feature 10h) and Figure 398; Identify Controller MNTMT, MXTMT, HCTMA (Figure 312)
NVMe power states and HCTM	The SMART / Health log carries a transition count and a total time in seconds for each of TMT1 and TMT2, and the counts only cover transitions caused by the host controlled thermal management feature. A drive with HCTM left at 0 may show zeros.	[S]	NVMe Base Specification 2.1, SMART / Health Information Log (Figure 206)
NVMe power states and HCTM	Hagiwara tested an industrial SSD with TMT1 80°C and TMT2 85°C. Throttling began at 80°C, writes dropped by up to 100 MB/s, and it ended near 75°C, repeating as a cycle.	[V]	Hagiwara Solutions tech blog (hagisol.com/techblog?p=635)
NVMe power states and HCTM	Some enterprise drives let the host pick the operating power: Solidigm D7-P5500 at 15, 20 or 25 W, and D7-PS1010 from 5 to 25 W.	[V]	NVM Express, Technology Power Features; Solidigm product pages
NVMe power states and HCTM	Some consumer drives list the same maximum power for PS0, PS1 and PS2, so stepping the power state would save nothing there (Samsung 990 PRO dump: 9.39 W for all three). The lab says “some drives” and gives no drive name in the sentence.	[F]	User-posted smartctl output, Fedora and Arch forums
Fewer dies	Firmware can limit how many dies work at once to stay inside a power budget. A Phison patent describes doing this when a hot-spot temperature is high. That throughput falls roughly in proportion to the dies in use is the lab’s model assumption.	[P]	US20130080680A1 (Phison)
Fewer dies	A multitoken power budget gave 22.85% more throughput than the prior state of the art under a power cap, with 3.8% performance overhead.	[paper]	You et al., IEEE TCAD 39(10), 2020, abstract
Fewer dies	A research add-in card cut power 15 to 16% by reducing active internal resources and lost 12 to 18% performance at its throttle point.	[paper]	Zhang, Shihab and Jung, HotStorage 2014. A lab card at 11 to 18 W, not a consumer M.2 drive.
Fewer dies	Not measured on any named consumer drive. The page says so on screen.		Vendor research report, section 10, item 6
Clock reduction	Phison and WD/SanDisk patents describe lowering the controller work frequency or the NAND interface rate in tiers.	[P]	US20130080680A1, US11016545
Clock reduction	Innodisk describes “conventional” schemes that lower the flash clock at 85°C.	[V]	Innodisk white paper
Clock reduction	Not measured on any named consumer drive. The page says so on screen.		Vendor research report, section 10, item 6
Critical shutdown	At launch, Phison E26 drives (Corsair MP700, Crucial T700, Seagate FireCuda 540, Gigabyte Aorus Gen5 10000, Adata Legend 970) shut down instead of throttling. Phison made firmware changes after the reports, and firmware 22.1 made the MP700 throttle.	[R]	TechPowerUp MP700 review; Phoronix; ComputerBase; Phison statement relayed by FPS Review and HotHardware
Critical shutdown	The bare MP700 stopped after 86 s of reads or 55 s of writes and returned only after a power cycle (TechPowerUp, via FPS Review). The lab quotes only what the TechPowerUp page itself says.	[R]	TechPowerUp, Corsair MP700 2 TB
Critical shutdown	ExtremeTech, relaying ComputerBase, reported an early Crucial T700 throttling to about 102MB/s writes at 86°C in CrystalDiskMark without crashing. The lab says “about 102MB/s” and names ExtremeTech as the relay, because it did not read ComputerBase’s own page.	[R]	ExtremeTech, Corsair and Crucial Gen5 SSDs shutting down or severely throttling
Critical shutdown	TechPowerUp reported a Team Group GA Pro (not a Phison E26 drive) that crashed after a few minutes in a throttled state. This is the only source for the lab’s “crash while throttled” outcome, and the timings in the lab are model values.	[R]	TechPowerUp, Team Group GA Pro 2 TB review
Critical shutdown	Phoronix saw errors “always within three minutes” with the drive reporting about 87°C, and EXT4 then mounted read-only. That is the file system reacting to I/O errors. No source describes a drive that makes itself read-only because of heat.	[R]	Phoronix, PCIe 5.0 NVMe SSD failures
Critical shutdown	Crucial lists throttling from 81°C and protective shutdown at 90°C for the T700 and T705.	[V]	Crucial T700/T705 specifications as relayed by Guru3D, StorageReview
Critical shutdown	The NVMe composite temperature threshold CCTEMP is a critical overheating condition whose consequences may include automatic device shutdown, extreme performance throttling, data loss or damage. It does not require any of them.	[S]	NVMe Base Specification 2.1, Identify Controller, CCTEMP
Critical shutdown	A controller reset or CSTS=0xffffffff in a Linux log has also appeared with a kernel message asking whether the device has a faulty power-saving mode, so a log line alone never shows overheating.		Look-alike named in the failure section; Linux nvme timeout messages
Host link throttling	Linux 6.13 adds a PCIe bandwidth controller (bwctrl) and a PCIe cooling device that can lower the root port link speed. Width cannot be changed before PCIe 6.0 L0p.	[S]	LWN, PCIe bandwidth controller; Linux kernel thermal documentation
Host link throttling	The kernel’s step_wise governor works from the trend as well as the trip point. Above a trip point it uses a higher cooling state while the temperature is rising and a lower one while it is dropping. Below a trip point it does nothing while rising and uses the minimum cooling state while dropping.	[S]	docs.kernel.org, thermal governors
Link width	No source describes any SSD lowering link width (x4 to x2) for heat. The lab does not model it.		Vendor research report, section 10, item 4

Cross-cutting behavior

Claim	Grade	Source
In one measured example, a Hagiwara SN8E industrial SSD, throttling began at the 80°C TMT1 threshold and ended near 75°C.	[V]	Sources above
With TMT1 set below the temperature a throttled drive settles at, the lab’s model stays throttled. This is the lab’s own model result and no source gives a stuck drive. Hagiwara’s 80/85°C test releases about 5°C below TMT1 and cycles.	[MODEL]	Hagiwara tech blog (hagisol.com/techblog?p=635)
Writes usually throttle harder than reads: Samsung 9100 Pro bare, writes about 1 GB/s against reads 2 to 5 GB/s. Not universal: Crucial P5 throttled on reads first, and Crucial T710 with a large heatsink stopped write throttling while reads still dipped.	[R]	TechPowerUp reviews of the 9100 Pro, Crucial P5 and Crucial T710
Reported temperature is not the hottest spot. The NVMe composite is implementation specific. TechPowerUp measured 106°C at the 9100 Pro surface against a lower SMART value. ATP says a sensor far from the hot part makes throttling engage late.	[S] [R] [V]	NVMe Base Specification 2.1, Composite Temperature; TechPowerUp 9100 Pro review; ATP blog
A drive that is throttled can still overshoot: the temperature kept climbing after throttling began until shutdown (76°C throttle, 86°C shutdown) on a computational SSD.	[paper]	Waltz, ACM TOS 2026 (arXiv 2509.05365)
Facebook’s fleet study found platforms that throttled aggressively had lower failure rates than platforms that rarely throttled.	[paper]	Meza et al., SIGMETRICS 2015
Sean’s own published numbers: the IronWolf 525 began to throttle after roughly 500GB of writes at 83°C; the KC3000 reported 71°C after about 500GB and engaged its throttle then.		Sean’s Seagate IronWolf 525 and Kingston KC3000 reviews

Model values (no published source; every one is labeled “model value” in the page)

Value	Default	Reason
Thermal resistances (two nodes)	Controller node to ambient: bare 22, shield 15.8, heatsink 14.1, heatsink plus fan 7.3°C/W. NAND and board node to ambient: 10, 7.2, 6.4, 3.3°C/W. Between the two nodes: 14°C/W in every case.	Fitted so a Gen4 drive under a sustained write settles near 80°C controller and 73°C NAND bare at 30°C ambient (the common pattern of a Gen4 drive near the 80s without a heatsink) and a Gen5 drive crosses 85°C in about 2.5 minutes. The other three scale the resistances. No paper publishes C/W for an M.2 2280.
Controller time constant	5 to 30 s	Analysis Tech multi-time-constant example; two-node shape
Board and NAND time constant	3 to 15 min, longer with a heavy heatsink	Observed times to throttle of about 10 min (Waltz) to 18 to 24 min (HotStorage 2014); WD ITherm 2019 uses Foster RC terms for M.2
Heat from the controller and NAND	Gen4 about 7W total under a sequential write, Gen5 about 10W; writes cost more than reads	Press and review figures: SM2508 about 3 to 3.5 W controller, E26 about 5 W; TechPowerUp measured Gen5 drives from 9 to 14 W peak. Order of magnitude only.
Power per active NAND die	about 0.3 to 0.66 W while an array operation runs	Derived from ONFI 5.2 ICC maxima at about 3.3 V; a maximum, not typical
Peak power	NAND write power x 1.88 (a simulated 3.52W peak over a 1.87W average), plus the controller at its average: Gen5 sustained write 10.5W average, 15.6W peak	A published SSD simulation (SSDExplorer, Zuolo et al.) of a 128-die mid-1x-nm TLC subsystem (16 channels x 8 dies), random 25% programs and 75% reads in 4kB, driven by per-chip currents measured on a 4-chip TLC board (Toggle 400MT/s) with a sense resistor and an oscilloscope. The 1.87W and 3.52W and the suspend ratios are simulation results, not a measurement of the 128-die subsystem. The lab applies the ratio to every write workload, which is its assumption. For a per-chip NAND power model, see Mohan et al., “Modeling Power Consumption of NAND Flash Memories Using FlashPower,” IEEE TCAD 32(7), 2013, doi:10.1109/TCAD.2013.2249557.
Program suspend	One 200us window on the upper-page programs: peak -11.5%, average -3.5%, run time +5%, upper-page latency +16%. Two windows: -13%, -5%, +10%, +27%	The same simulation. The mechanism, program suspend, is described by Wu and He for letting reads go ahead of programs and erases (“Reducing SSD Read Latency via NAND Flash Program and Erase Suspension,” FAST 2012). For NAND designed to cut program power, see Sako et al., ISSCC 2015, doi:10.1109/ISSCC.2015.7062959.
DRAM power	One DRAM chip draws 0.19W powered and ready, 0.68W reading flat out and 0.91W writing flat out; the extra draw while serving accesses is a few tens of milliwatts, so throttling leaves the standby floor	Micron 16Gb DDR5 SDRAM die data sheet (Rev. C, 01/2023), Table 6, DDR5-4800 x16 maximum currents at VDD = VDDQ = 1.1V: IDD3N 142mA + IDDQ3N 31mA, IDD4R 530mA + IDDQ4R 92mA, IDD4W 479mA + IDDQ4W 348mA. The power is current times 1.1V, and the datasheet gives maximums under worst-case conditions, so these are ceilings. The part is a DDR5 stand-in: consumer drives mostly carry LPDDR4 or DDR4. The 5% burst share is the lab’s assumption; Park et al., “A comprehensive study of energy efficiency and performance of flash-based SSD,” Journal of Systems Architecture 57(4), 2011, give the sizes of the standby and access draws (about a hundred milliwatts per low-power chip, a few hundred for ordinary SDRAM, a few tens of milliwatts more during accesses). The lab puts the chip’s heat on the controller node.
Link ceilings	Gen3 x4 about 3.5 GB/s, Gen4 x4 about 7 GB/s, Gen5 x4 about 14 GB/s (usable, after overhead)	PCIe raw rates less encoding and protocol overhead; round numbers
Cooling thresholds for each preset	See throttle-presets.ts; each says whether it comes from a named source or is a model value	Evidence-labeled in the preset’s source field
Level 8 of the 950 PRO table	6.25%	Midpoint between level 7 (12.5%) and a floor; the white paper gives no figure

Read your drive: commands

Every command and option the lab prints was checked against a manual page or specification on 2026-10-04.

Command	Checked against
nvme id-ctrl /dev/nvmeX and nvme id ctrl /dev/nvmeX	nvme-cli v2.11 and master manual pages: the 2.x name is id-ctrl; the 3.x name is id ctrl and the bare id-ctrl is a deprecated alias. The temperature fields are WCTEMP, CCTEMP, MNTMT, MXTMT and HCTMA in NVMe Base Specification 2.1, Figure 312.
nvme smart-log /dev/nvmeX and nvme log smart /dev/nvmeX	nvme-cli master: smart-log is a deprecated alias for log smart and prints a warning. The log fields are NVMe Base Specification 2.1, Figure 206.
nvme get-feature /dev/nvmeX -f 0x10 -H (nvme-cli 2.x) and nvme get-feature /dev/nvmeX -f 0x10 (3.x)	nvme-cli v2.11 manual page lists -H, --human-readable; the master page lists no -H. Feature 0x10 is named hctm in the master page; the v2.11 page does not name it. The 3.x value is decoded by hand: TMT1 in bits 31:16, TMT2 in bits 15:0, in kelvins (NVMe Base Specification 2.1, Figure 398). Read-only: Get Features changes nothing.
smartctl -a /dev/nvmeX	smartctl(8): -a prints all SMART information; the NVMe device forms are /dev/nvme[0-9] and /dev/nvme[0-9]n[1-9]
hwmon: temp1_input, temp1_max, temp1_crit, temp1_alarm	Linux drivers/nvme/host/hwmon.c: temp1_input is the composite temperature from the SMART log, temp1_max comes from the NVMe Temperature Threshold feature (04h, over-temperature), temp1_crit is CCTEMP, temp1_alarm follows the critical-warning temperature bit. Whether the kernel takes any action on those values is not confirmed, so the lab does not claim it.
Get-PhysicalDisk | Get-StorageReliabilityCounter	Microsoft Learn, Get-StorageReliabilityCounter: lists Temperature and TemperatureMax among the reliability counters. The page does not say what TemperatureMax measures, so the lab does not say either.

How the model works

The drive has two thermal nodes: the controller, which heats and cools in seconds, and the NAND and board, which take minutes. Each node has a heat capacity, a thermal resistance to the air, and a coupling resistance to the other node. Every half second the model adds the power each node draws, takes away what it sheds, and steps the temperature forward. The method is the lumped thermal network Skadron and colleagues used for processors.

Power comes from what the drive is doing. The controller has a base draw, a draw that grows with activity, and a PHY draw that depends on the PCIe generation. The DRAM chip beside it draws a standby floor all the time and a little more while it serves lookups. Each busy NAND die adds its share, and a write costs more than a read. Throttling works because it makes the drive do less: the same state that sets the speed sets the heat.

Heat follows the average power, but the power supply sees the peaks. One published simulation of 128 TLC dies (16 channels of 8), driven by chip currents measured on real TLC NAND, under a random workload of 25% programs averaged 1.87W and peaked at 3.52W, so the lab draws the NAND’s write power as peaking 1.88 times its average. A bare Gen5 drive under a sustained write averages 10.5W and peaks at 15.6W. Peaks last microseconds, so meters that average over milliseconds read lower, and the peak is worked out from each half-second step rather than drawn.

The Program suspend setting pauses the upper-page programs, the most power-hungry ones, so their peaks do not stack. In the same simulation one 200µs window took 11.5% off the peak for 5% more time and 16% longer upper-page programs; two windows took 13% off for 10% more time and 27% longer upper-page programs. One window takes the peak to 14.2W. It is a power cap, not a thermal throttle: the average, and so the heat, barely moves.

The Onboard DRAM setting adds the chip’s power to the controller side, where the chip sits on the board. The figures come from one vendor datasheet, Micron’s 16Gb DDR5 SDRAM die, at 1.1V and the maximum currents it lists: a chip that is powered and ready draws 0.19W, one reading flat out draws 0.68W, and one writing flat out draws 0.91W. A drive bursts to its DRAM for only a small share of the time, so the lab adds 5% of the gap between burst and standby at full speed, a few tens of milliwatts, which is the size a published energy study of SSDs gives for the extra draw while DRAM serves accesses. Standby does not fall when the drive throttles, so a throttled drive keeps that floor. Choosing no DRAM takes about 0.2W off the controller side: about 2% of a Gen5 drive writing flat out, and about a fifth of an idle one.

Each way of throttling is one small function that runs once a second. It reads the temperature the drive’s sensor sees, which is a blend of the two nodes you choose, and changes one thing: a level, the time spent waiting between commands, the link generation, the power state, the number of busy dies, or the clock. A drive that reaches its shutdown temperature stops answering. A drive with no throttle at all has no protection.

Each way of throttling draws its own shape on the speed chart: a staircase, a smooth ramp, a step between two link speeds, two levels, steps of one die, a uniform drop, a fall to zero, and a step the host chose. Each failure is the same chart with different settings: a smaller gap between entry and release, a lower threshold, or a sensor that sits farther from the hot part.

Terms

• controller. The controller is the chip that runs the drive: it talks to the computer, manages the NAND and decides when to throttle.
• PHY. The PHY is the part of the controller that drives the PCIe link, sending and receiving the signals. It draws more power at higher link speeds.
• NAND die. A NAND die is one piece of flash memory. A drive spreads its work across many dies, and every die that is busy draws power.
• watt. A watt is one joule of energy per second. A drive that draws 10W turns 10 joules into heat every second.
• thermal resistance. Thermal resistance says how many degrees hotter a part gets for each watt it must shed. A heatsink lowers it.
• time constant. A time constant is how long a part takes to cover about two thirds of the way to its new temperature. The controller takes seconds. The board and NAND take minutes.
• ambient. Ambient is the temperature of the air around the drive.
• WCTEMP. WCTEMP is the Warning Composite Temperature Threshold in the drive’s Identify Controller data. The SMART log counts the minutes spent above it.
• CCTEMP. CCTEMP is the Critical Composite Temperature Threshold. The NVMe specification says it may lead to shutdown, extreme throttling or damage, and requires none of them.
• composite temperature. The composite temperature is the single number an NVMe drive reports for the whole drive. How it is worked out, from which sensors, is up to the maker.
• throttle level. A throttle level is one step in a table of speed caps. Level 0 is full speed.
• hysteresis. Hysteresis is a gap between the temperature where a throttle starts and the lower one where it stops, so it does not flicker on and off.
• I/O pacing. Pacing means waiting between commands so the NAND does less work each second.
• latency. Latency is how long one command takes to finish.
• PCIe generation. A PCIe generation sets the speed of the link. A Gen5 x4 link carries about 14GB/s, Gen4 about 7GB/s and Gen3 about 3.5GB/s, after overhead.
• link-state throttling. Link-State Thermal Throttling is Phison’s name for lowering the controller’s own PCIe link rate to cool down. It was introduced for the E26 in firmware 22.1.
• power state. An NVMe power state caps how much power the drive may use. A lower state can cost speed, though some drives list the same power for several states.
• HCTM. Host Controlled Thermal Management lets the host set two temperatures at which the drive throttles lightly and then heavily. It is an optional NVMe feature.
• TMT1 and TMT2. TMT1 and TMT2 are the two Thermal Management Temperatures. The SMART log counts how often and for how long the drive crossed each.
• power budget. A power budget limits how much power the drive may use at once. Firmware can meet it by running fewer NAND operations at the same time.
• die parallelism. Die parallelism is how many dies work at the same time. More dies means more speed and more power.
• clock scaling. Clock scaling lowers the speed of a clock to save power. Here it means the controller clock or the NAND interface rate.
• critical shutdown. A critical shutdown is the drive stopping itself, with no further answers to the computer, to avoid damage.
• power cycle. A power cycle removes power from the drive and restores it, which clears a latched shutdown.
• PCIe bandwidth controller. The PCIe bandwidth controller is a Linux 6.13 feature that lets the operating system lower a link’s speed as a cooling action.
• overshoot. Overshoot is the temperature rising past the trigger because heat already stored keeps flowing after the throttle acts.
• controller reset. A controller reset restarts the drive’s controller while the computer keeps running. The drive is away until it finishes.

Checked and not shipped

• What heat does to NAND. The Arrhenius retention table (HeatWatch, Luo et al., HPCA 2018), the warm-programming findings and the JEDEC JESD218 client class were in the old walkthrough and are no longer on the page, so their rows are gone. Not claimed.
• LSTT visible to the host. No review or log shows a Gen5 to Gen4 change in lspci, HWiNFO or GPU-Z on a throttling E26 drive. Not claimed.
• Phison TT1, TT2, TT3 temperature steps. Never published for any Phison controller. Only the 950 PRO has a published table.
• Link width (x4 to x2) throttling. No source on any SSD.
• Heat-caused read-only mode. No source. The read-only cases found were file systems reacting to I/O errors.
• Phison’s “10x to 20x faster than previous throttling.” Marketing claim, not independently measured. Not used.
• Rossmann Group claims about emergency power-down corrupting FTL metadata. A data-recovery vendor’s unverified claim. Not used.
• Samsung DTG step tables for the 970, 980, 990 and 9100. Not published. Only the 950 PRO is quoted.
• Whether the kernel acts on the NVMe temperature critical warning. Not confirmed. Not claimed.
• Windows Settings storage page and vendor toolboxes as places to read a drive temperature. Not checked against vendor documentation, so not named.
• Microsoft SYSTOR 2016 (Narayanan et al.) “surprisingly small effect” snippet. Could not be tied to the paper. Not used.
• Cold-temperature self-heating drives. No citable source. Not claimed.