16 GB DDR4-2666 PC4-21300 ECC registered RDIMM, 1Rx4 on eighteen 8 Gb x4 packages, 1.2 V, 288-pin; twelve ordering codes under the M393A2K40BB2, M393A2K40CB2 and M393A2K40DB2 revisions
| Capacity | 16 GB |
|---|---|
| Memory Type | DDR4 |
| Speed | 2666 MT/s |
| Form Factor | RDIMM |
| Rank | 1Rx4 |
| ECC | ECC |
| Voltage | 1.2 V |
| Module Height | Standard |
| Condition | New · Used |
Samsung 16 GB DDR4-2666 ECC registered RDIMM at standard height, one rank of eighteen x4 devices on a 288-pin board at 1.2 V. PC4-21300 at CAS latency 19, 14.25 ns at the 1,333 MHz clock, and 21.3 GB/s peak per 64-bit channel.
This is the fastest bin a mainstream DDR4 server will still hold at two modules per channel, and the build that gets closest to holding it is the single-rank one. That is only half the picture. A single rank also means one set of open banks per channel rather than four, and at 2,666 the balance between clock and bank concurrency is what separates a well-matched machine from a merely fast one.
Three die revisions carry it here — M393A2K40BB2, M393A2K40CB2 and M393A2K40DB2 — under four, six and two ordering codes respectively. Twelve codes in one table at the foot of the page, one set of specifications above them.
| Part number | M393A2K40BB2-CTD, M393A2K40CB2-CTD and M393A2K40DB2-CTD — three revisions of one module, with the twelve ordering codes listed under Ordering codes |
|---|---|
| Capacity | 16 GB per module, single module |
| Organisation | 1Rx4 — one rank of eighteen x4 devices at 8 Gb each, eighteen packages carrying 128 Gb of data plus the eight check bits that ride with every 64-bit transfer; 2G x 72 at module level |
| Speed | DDR4-2666, PC4-21300, 2,666 MT/s, 21.3 GB/s peak per 64-bit channel |
| Timings | CL19 at 1,333 MHz, 14.25 ns; JEDEC 2666V bin |
| Voltage | 1.2 V nominal |
| Form factor | 288-pin RDIMM at standard height, 31.25 mm, bare PCB with no heatspreader |
| Register | On-module register buffers the command and address bus, so the board has to accept registered memory |
| Error handling | ECC across the 64-bit data path with x4 granularity, which is what allows a failing device to be retired instead of tolerated |
| Manufacture | Samsung; assembly site and date code vary by production run and are printed on the module rather than fixed by the part number |
| Condition | New and tested pulls, both stocked |
Compatibility is decided inside the part number, not on the label. The label adds production detail, and reading the two as a single string is how capacity gets confused with rank.
| Printed as | Means | Why it matters at purchase |
|---|---|---|
| M393 | Standard-height registered module | The 31.25 mm board; the low-profile version of this capacity carries M392 |
| 2K | Sixteen-gigabyte class on 8 Gb devices | Fixes capacity and device density together, which is why eighteen packages reach sixteen gigabytes |
| 40 | x4 device width | The width that gives chipkill granularity, and the reason a rank is eighteen devices wide rather than nine |
| 1Rx4 as printed on most listings | Single rank, x4 devices | One rank against a channel budget, so two modules per channel stay well inside the usual table |
| PC4-21300V | DDR4-2666, bin V, which is CL19 | The letter states the timing grade; a listing that quotes only the clock leaves latency unstated |
| -R | Registered, not unbuffered | A different electrical interface from an unbuffered module of the same capacity, and it will not train in that slot |
| BB2 / CB2 / DB2 | Successive die revisions of one design | Records which silicon generation built the assembly; all three carry the same organisation, clock and voltage |
| CTD | The 2,666 speed block | Neighbouring blocks in this family are other clocks, so the block letter needs checking rather than assuming |
| 6Q … 8Y | Ordering suffix, two characters each | What the purchase order quotes; each revision has its own tail set, so a tail alone does not name a build |
| YYWW | Date code, year and week | The only service history a pull carries, and the first field to read when judging remaining life |
A rank answers together. Every device in it opens the same row at the same time, so the controller sees sixteen banks per rank, not sixteen times eighteen. Two 2Rx4 modules in a channel present sixty-four open banks; one of these presents sixteen. The clock does not change that ratio and neither does the capacity, and the ratio is what sets how many unrelated streams a channel can keep in flight before it starts waiting on a row to close.
For work that walks memory in order — backup, streaming analytics, a large copy, media pipelines — banks are beside the point and the 2,666 clock is collected in full. For work that chases pointers through a working set larger than the last-level cache, a four-rank channel at 2,400 will often finish first despite a lower paper figure, because it has four times the open rows to hide latency behind. Both statements are true at once, which is why the grade alone does not pick the part.
Where this module sits is the middle: the highest clock a two-module channel will still hold, with fewer ranks to keep coherent once the machine is full. It is the right part for streaming-heavy estates, and for builds where two modules per channel is the plan and clock is the only lever left.
Two of these are two ranks against a budget of four, so most boards keep the clock at full population. The same two slots filled with 2Rx4 modules are four ranks of thirty-six packages, and that is where conservative firmware starts easing back to 2,400 without telling anyone. The difference is not visible in any specification; it appears in one line of output, the configured speed reported by the operating system, which is worth reading once per build rather than assuming.
If the platform reports 2,400 after assembly, the memory is running as designed and nothing is broken — the board decided the population did not justify the higher clock. In that case the money spent on the 2,666 grade is not recovered, and a 2,400 part would have produced the same machine for less. That is the honest test, and it costs one command at build time instead of a purchase order in the wrong grade.
Absolute latency across the common DDR4 bins runs 14.06 ns at 2,133 CL15, 14.17 at 2,400 CL17, 14.25 here, and 13.75 at 3,200 CL22. The middle two are within a tenth of a nanosecond of each other, so latency is not what separates 2,400 from 2,666; bandwidth is, at 19.2 against 21.3 GB/s. The 3,200 grade is the outlier that recovers latency because its CAS count rises more slowly than its clock, and it is also the grade most likely to be given up when a board is full.
Refresh overhead rises with capacity per rank rather than with clock, so the single-rank layout is mildly kinder here than a doubled-up rank would be: fewer devices to refresh per command, and a shorter tRFC window in which the channel is unavailable. At sixteen gigabytes the effect is small, and it is a tie-breaker rather than a reason to buy.
The channel takes the lower of the module rating and the platform rating, and on these three the module is the one that runs at label speed or below it:
None of the three needs configuring, and none is harmed by the difference in rating. The part number describes the module; what the machine reports afterwards is firmware's decision.
Twelve codes carry this module: four under M393A2K40BB2, six under M393A2K40CB2 and two under M393A2K40DB2. The family code fixes height class, density, device width and revision, CTD fixes the 2,666 speed block, and the last two characters name the build.
The three families do not share a tail set. BB2 is stocked under 6Q, 7Q and 7Y; CB2 under 6Y, 7Q, 7Y, 8Q and 8Y; DB2 under 7Y alone. Only 7Y appears in all three, and the base code CTD is the other thing every family has by definition. So a tail can be carried to another revision only after checking that it was issued there — keeping the tail and swapping the letters can name a build that never existed.
On the revisions themselves: all three are sixteen gigabytes of DDR4-2666 at 1.2 V in the 1R x 4 organisation of eighteen 8 Gb devices, and the letters record successive silicon generations of one raw card rather than any change in capability. Samsung lists the DB2 revision as end of life on its own product page; the earlier two are held here under the same cover and run beside it in a channel without distinction.
| Code | Build | Build tail | Condition | Quote |
|---|---|---|---|---|
| M393A2K40BB2 | BB2 catalogue | Family code without a speed block; the BB2 build listed here is the PC4-21300 / 2,666 MT/s one | New / used | Enquire |
| M393A2K40BB2-CTD | BB2 | Base code, no suffix appended | New / used | Enquire |
| M393A2K40BB2-CTD6Q | BB2 | 6Q | New / used | Enquire |
| M393A2K40BB2-CTD7Q | BB2 | 7Q | New / used | Enquire |
| M393A2K40BB2-CTD7Y | BB2 | 7Y | New / used | Enquire |
| M393A2K40CB2 | CB2 catalogue | Family code without a speed block; the CB2 build listed here is the PC4-21300 / 2,666 MT/s one | New / used | Enquire |
| M393A2K40CB2-CTD | CB2 | Base code, no suffix appended | New / used | Enquire |
| M393A2K40CB2-CTD6Y | CB2 | 6Y | New / used | Enquire |
| M393A2K40CB2-CTD7Q | CB2 | 7Q | New / used | Enquire |
| M393A2K40CB2-CTD7Y | CB2 | 7Y | New / used | Enquire |
| M393A2K40CB2-CTD8Q | CB2 | 8Q | New / used | Enquire |
| M393A2K40CB2-CTD8Y | CB2 | 8Y | New / used | Enquire |
| M393A2K40DB2 | DB2 catalogue | Family code without a speed block; the DB2 build listed here is the PC4-21300 / 2,666 MT/s one | New / used | Enquire |
| M393A2K40DB2-CTD | DB2 | Base code, no suffix appended | New / used | Enquire |
| M393A2K40DB2-CTD7Y | DB2 | 7Y — the one tail issued in all three revisions | New / used | Enquire |
Nothing here is priced because the line is quoted per order: two conditions, twelve codes and a per-piece number that moves with quantity. Send the code, the quantity, the condition and the platform, and the answer comes back the same working day.
Modules ship with one year of cover counted from the dispatch date, new stock and tested pull alike, and the cover follows the ordering code and the batch rather than the invoice. Every pull passes a matched-set screen before it is packed, and the screening record goes into the box with the module. Stock on the shelf is dispatched in three to seven working days with tracking, and the smallest order accepted is two modules.
Tell us who you are — we reply within one working day with a firm quote, stock status and lead time. Volume discounts available.
16 GB DDR4-2933 PC4-23400 ECC registered RDIMM, 2Rx8 on eighteen 8 Gb x8 packages, 1.2 V, 288-pin; six ordering codes under M393A2K43CB2-CVF with BY, CO, BQ, C0 and GQ tails
16 GB DDR4-2400 PC4-19200 ECC registered RDIMM, 2Rx8 on eighteen 8 Gb x8 packages, 1.2 V, 288-pin; eighteen ordering codes under the M393A2K43BB1 and M393A2K43CB1 revisions
16 GB DDR4-2133 PC4-17000 ECC registered RDIMM, 2Rx8 on eighteen 8 Gb x8 packages, 1.2 V, 288-pin; nine ordering codes from M393A2K43BB1-CPB through the CPBQ tail
16 GB DDR4-2666 PC4-21300 ECC registered RDIMM, 2Rx8 on eighteen 8 Gb x8 packages, 1.2 V, 288-pin; six ordering codes under M393A2K43BB1-CTD with 6Q, 6Y, 7Q, 7Y and 8Q tails