The fastest mass-market storage device has been the NVMe module and it has also been the quickest to fail. It will stop with little or no advance warning. In other words, there is no noise and the machine does not slow over time; a working machine at one moment, and the next a boot failure stands in place of where the drive should have been. M.2 modules and PCIe cards from every maker come in from home machines, gaming builds and the research and engineering offices near the A11.
Every nvme job is diagnosed free. You get a fixed quote in writing before a single screwdriver is picked up.
In most cases you only pay if we recover your data. Electronic, mechanical, chip-level and physical work is charged at 50% upfront. Ransomware, forensic, BitLocker and CCTV/DVR jobs are charged 100% upfront on acceptance of our fixed quotation. Please refer to the data recovery cost page for full details on costs.
The first thing done on each nvme job is matching the symptom to the fault; after twenty plus years, these thirty account for nearly everything that comes through the door.
Controller or firmware trouble cleared the bus overnight. Behind that silence the data is intact and just waiting for some way to be managed again.
A dead controller, bad power stage or corrupt boot code are all possible causes, but dead memory would have to be the least likely of the four. The chances of recovering the data are good; the module simply does not know how to tell you this.
Translator failure results in a module answering with a capacity found on no price list. Service mode repair fixes it, and it is generally well marked.
The system module stopped talking mid sentence. Once removed from the machine and taken to the bench, it will give up its files without much protest.
Bare modules may throttle, disappear under load, and scramble their tables over repeated hot spells. Imaging them cold solves a surprisingly large number.
Whether it was a Windows update or a vendor supplied utility, the end result will be the same: a module that cannot even tell us what it calls itself. It gets reflashed here, or else read around completely.
Flash does not keep a log book and it never sends you a postcard. Perfect service from beginning to end, right up until it collapses, is the usual story, which is why backups matter more here than on any other item.
Thin boards lose to upgrade levers and stiff envelopes. When there are track repairs, they cover the small cracks. Chip-off covers everything else.
Surge damage to the power circuits is repaired by replacing damaged components at board level before anyone tries to read anything. The order is the whole job.
Background wiping of freed blocks starts after deletion and only stops when power is turned off. If you pull the plug, the clock simply stops where it is.
WD's top-of-the-line drive came out with firmware that hung and disappeared under certain loads until it was fixed. Unfixed units continue to keep our booking sheet interesting.
The 980 and 990 generation used up NAND life at a rate that needed an update to correct. Units that were never updated still arrive with their endurance largely gone.
Kingston's NV2 includes whatever controller and flash were available that quarter. Two drives will have the same name but be very different inside, and what is actually inside decides the route we take.
Steam Deck and other handheld systems build their small modules into tight, enclosed spaces where there is no room for heat to escape. Small format, but as far as recovery goes, these fit the typical mould.
A DRAM-less module borrows memory from its host computer to store the map of its contents. If a single interaction goes wrong, what was borrowed gets corrupted. Service mode corrects the ledger.
When a fast module runs bare without cooling, it ages quickly due to heat. The first failures show as dropouts, followed later by complete disappearance, and the dropout stage is when you want the module imaged.
Sustained shader compilation and continuous cache thrashing put roughly a year of wear on a gaming module each season. Stuttering is the early warning sign and the disappearance is the deadline.
Many budget designs have been known to corrupt their mapping tables after going to sleep or after a crash, because of how they use borrowed host memory. This type of corruption is well documented and has a well documented way out.
Budget makers beat their pseudo-SLC area until it goes bad and the drive either stalls or simply bricks. Usually the slower flash behind it still works fine.
Over tight heatsink screws and mangled thermal pads bend the board and crack solder joints. It gets reworked under the microscope, then imaged.
After some BIOS updates, or with a weak riser card, the module and the slot can't come to terms, so no module shows up. Easy to prove on a known-good lab socket instead of trying to find the problem by swapping cables.
When just one die fails, the drive shrinks, errors appear across its stripes, or it locks in read-only mode. Imaging works around the die that has gone and lets its neighbours carry the data out.
When power is lost during housekeeping, the map gets partly rewritten and the drive mounts, loops and sulks in turn. Service-mode reconstruction puts everything back to normal.
Counterfeit units are recovered for what they are, not what the label says they are, and the label fools nobody here.
Cracked socket latches and lost PCIe lanes leave a module limping along at a fraction of its width and then gone altogether. Imaged on an intact lab socket.
Reformatting an NVMe drive to a new sector layout is as easy as issuing one command, and the mapping of all sectors is thrown away completely. It is fast, quiet, and leaves what looks like a blank drive. Recovery then happens by accessing the raw memory and reconstructing how the sectors were arranged.
Both commands are intended to be irreversible and each is very good at it. Once a sanitize has completed there will be nothing left behind. If a sanitize was interrupted, many times there will still be a great deal left, and the free diagnostic will tell you which happened instead of guessing.
There are known cases in certain combinations of modules and platforms of aggressive low power states causing the drive to drop out, most notably on Linux. The drive falls asleep and does not wake up, and it looks terminal. Often the module is fine and the tables just need rebuilding after the crash it caused.
A full drive cannot do its own housekeeping, because in order to move data from one place to another it must have somewhere to put the data. Write performance slows down dramatically, errors start to build up, and some drives go into read-only mode. That read-only mode is a pain to live with, but it helps when imaging.
The same M.2 socket on two machines can carry very different signals, so a module that fits may not be connected to anything it can talk to. Nothing shows up, nothing is broken, and hours get spent on a fault that doesn't exist. It is worth checking before writing off the drive.
An M.2 module communicates to the processor via PCI Express (with no mechanical components in between), and this is why an M.2 is faster, and also why its failure can be very abrupt. There is no "limp" or "noisy" phase of failure that would allow the user to know something was wrong. A dense firmware controls how data is stored on stacked NAND flash; and once either the controller or that firmware loses its footing, the module will stop responding. However, the data in the memory that sits behind the silent controller remains perfectly intact. Getting to it comes down to either the manufacturer's service mode where the controller may still respond, or accessing the individual chips with a rebuilt translation layer where it will not. In neither case does access improve after you have restarted the system another dozen times at your kitchen table, although it is certainly understandable that one might wish to keep trying.
Modules in the 2280, 2242 or 2230 length (and also PCIe add-in cards), from Gen3 to Gen5, made by Samsung, WD, Crucial, Kingston, SK hynix, Kioxia, Sabrent and Solidigm, can all be worked on so long as they unplug from their slot and can be posted to us. The one exception has remained soldered NVMe, which we decline. Heat needs a mention of its own. If you have a bare stick with no heatsink, it will throttle, then fall off the bus altogether, and over repeated episodes it can scramble its own mapping tables; every summer small form factor systems and slim laptops send us a run of exactly that. Whatever caused your failure, the free assessment identifies the cause and gives you a price before anything is spent.
Controller, firmware and chip surgery at PCIe speed, with the bench set up so the work never waits on equipment:
Silent modules are brought up in the manufacturer's own service mode, the firmware is corrected and the translator rebuilt, and then the image is taken without ever gambling with the memory.
Modules that stall or wander are captured under hardware enforced timeouts, using retry behaviour that is not accessible to software tools.
From 2230 stubs to full length 2280s, add-in cards to enterprise U.2, everything is connected natively, with active cooling for modules that start to fade when they warm up.
When service mode can't reach it, each package is removed and read one by one. A longer way around to the same end.
Interleave, XOR and ECC are worked out from the dumps until the mapping stands up again, and the volume then stands on it.
Burned power stages are built back component by component, and cracked boards are reworked under the scope before any attempt to image.
Reputations on this page have been earned the hard way. The early SN850X firmware would wobble at high loads until WD fixed it; many of the 980 generation Samsung drives which never received an update still arrive with their endurance largely spent; with an NV2, what is inside the drive decides how we go about accessing its data; and Gen4 modules fitted without a heatsink slowly cook themselves. Everything from 2230 up to 2280, including add-in cards, U.2, and Gen3 through to Gen5, can be worked on as long as the module unplugs and can be posted. Soldered-down NVMe is the one refusal, and phones and tablets are the other.
Most items arrive at this bench by insured and tracked post. We recommend getting a quote first, either by telephone or using our free online diagnostic, and once you have received your quote, simply post the item. Post is the steadiest way to move a poorly drive, and a parcel sent from Norfolk typically arrives on the bench within a working day.
Is the drive still bolted inside a laptop, desktop, MacBook, iMac, server or CCTV / DVR recorder? Before sending, remove the hard disk drive (HDD) or solid state drive (SSD), as we do not accept whole units. From a DVR or recorder, send just the disk. From a server or RAID, take out the disks and clearly label each disk by bay number. A network attached storage (NAS) device may be sent as its disks if they can be removed; otherwise please send the NAS itself. There is only one type of drive we are unable to accept: storage soldered onto a motherboard (e.g. Apple Silicon Macs, extremely thin laptops). In such cases, if the storage cannot be taken out it cannot be accepted. If you are unsure whether you have a removable drive or not, call us before sending anything.
↓ Print the shipping & booking-in form (PDF)
Please mark the parcel for the attention of Cambridge Data Recovery. It takes about an hour and twenty minutes to travel from Norwich along the A11, then two minutes off the A14 at Junction 32. Otherwise you can post it using a tracked service and we will receive it the following working day. We will contact you as soon as we have your drive and it is booked onto our bench.
If you're unsure of what to include in the box, please call us on 0800 689 0668 before sealing it. Or, for a first view on what we can do, you can use our free online diagnostic.
Free diagnosis, one fixed figure written down, and in most cases you only pay if we recover your data. Start online or ring us.