Some of the optics in this range do not have part numbers at all. ACHU, AJHE, AJHF, AJHJ et AJHK are IBM feature codes: four characters that identify something you can order against a machine, not something with a catalogue number. Ask a distributor for an ACHU and you will get a blank look. Ask IBM’s configurator and it knows exactly what you mean.
That is the theme of this guide. Server and storage optics are not ordered the way switch optics are ordered. They are specified against an array, a host bus adapter or a server, by people whose job is storage rather than networking, through a configurator rather than a price list — and the part numbers reflect that at every turn. What follows is how the Lenovo, IBM, NetApp and Intel schemes read, why one vendor’s numbers are another vendor’s with the prefix swapped, the Fibre Channel rules that decide whether a link comes up, and the 53 server and storage platform optics Carritech holds, from 10 Gigabit to 800.
One number tells you what this range is: 40 of the 53 parts are Fibre Channel. That is three-quarters, and it is by a wide margin the highest proportion of any vendor group on this site — the Dell range runs about two-fifths and the Brocade range about a third. If you are buying storage platform optics, you are almost certainly buying for a SAN, where the rules are not Ethernet’s and where getting the generation wrong leaves a path missing from a multipath configuration rather than a link down on an uplink.
The four naming schemes, and why one shelf holds two of them
| Family | Example | Vendor | What it tells you |
|---|---|---|---|
| Four-character feature code | AJHF | IBM | Nothing. It is an orderable feature against a machine type, not a part. |
| Eight-character alphanumeric | 01KN795 | IBM and Lenovo | Nothing readable. The shared heritage scheme — both vendors use it. |
| Machine-type style | 8960-2301 | IBM | A machine type and model rather than a component number. |
| 4M27A / 4TC7 prefixed | 4M27A65416 | Lenovo | ThinkSystem option numbers. The trailing digits run in blocks — decoded below. |
| X-prefixed | X6601A-R6 | NetApp | X then a number, often with an -R6 revision suffix. |
| X-000nnn | X-000238 | NetApp | A rebadged Brocade number. See below — this is the most useful thing here. |
| 332-00nnn | 332-00351 | NetApp | An older NetApp numeric series. |
| E-prefixed descriptive | E25GSFP28LRX | Intel | Fully readable: speed, cage, port type. The best scheme in this range. |
The IBM and Lenovo overlap is not an accident and it is worth understanding, because it explains why one storage room holds both prefixes against identical hardware. Lenovo acquired IBM’s x86 server business, with the initial closing completing on 1 October 2014 and country-by-country closings following over the months after. IBM kept Power Systems and the storage business. So the eight-character scheme — 00MY768, 88Y6393, 98Y2176, 00RY190 — runs across both companies, because it was one company’s scheme when those numbers were issued. The later 4M27A et 4TC7 prefixes are Lenovo’s own, and the four-character feature codes are IBM’s own.
The practical reading: a number in the shared scheme tells you nothing about which vendor supports the host, so you have to know the machine. A ThinkSystem server is Lenovo’s; a Power server or a FlashSystem array is IBM’s; and a part number issued before October 2014 could be sitting in either.
Lenovo’s ThinkSystem numbers do encode something, and it is useful
This is the one piece of decoding available in the range, and it is worth knowing because it covers eight of the parts below.
The 16 Gigabit Fibre Channel block runs 4M27A65411, 4M27A65412, 4M27A65413 et 4M27A65414. The 32 Gigabit block runs 4M27A65416, 4M27A65417, 4M27A65418 et 4M27A65419. In both blocks the four parts run short-wave, short-wave, long-wave, long-wave — so the last digit tells you the optical interface and the block tells you the generation, with 15 left unallocated between them.
That is genuinely helpful when you are reconciling an inventory export, and it comes with the usual caution: it is a pattern in a sequential allocation, not a specification. It holds across these eight parts and we would not bet a purchase order on it holding across a block we have not seen. Read it as a sanity check, not a substitute for the part number.
NetApp’s X-000nnn numbers are Brocade numbers with the prefix swapped
This is the most useful finding in this article and it saves real time.
X-000212, X-000213, X-000238 et X-000239 are NetApp 32 Gigabit Fibre Channel optics. XBR-000212, XBR-000213, XBR-000238 et XBR-000239 are Brocade 32 Gigabit Fibre Channel optics. They are the same four modules. Swap XBR for X and the numbers are identical, and our own mapping table confirms the match independently: both sets resolve to the same two Carritech designs with the same split — the 212 et 213 parts short-wave on multimode, the 238 et 239 parts long-wave at 10 kilometres.
That is not a coincidence and it is not a transcription slip. NetApp’s SAN switching comes from Brocade, so the optics in a NetApp-badged fabric switch are Brocade optics with a NetApp order code on the invoice. Three things follow and all three are worth money:
- If a NetApp X-000nnn number returns nothing, search the Brocade number. It is the same optic and it is far better documented, because Brocade published support matrices for a living.
- The array side and the switch side of the same fabric can be the same module under two vendors’ numbers. An estate holding both is holding one optical design as two line items, which is exactly the duplication that a smarter stock strategy is about.
- The identification strings are still different. A merger of part numbers is not a merger of EEPROM contents, and a NetApp switch looking for a NetApp string will not be satisfied by a Brocade one. That is what transceiver coding is for, and it is why one optical design can serve both ends if it is coded at the point of deployment rather than at the point of order.
Our guide to Brocade compatible optics covers the other side of this, including the three companies that now own different parts of the Brocade catalogue, and it is the companion piece to this one.
The Fibre Channel rules that decide whether the link comes up
Three-quarters of this range is Fibre Channel, and Fibre Channel is not Ethernet. The standards come from the INCITS Fibre Channel Technical Committee — known as INCITS/T11 until January 2022, and still called T11 by most of the industry — which is “responsible for the development of the Fibre Channel (FC) standards”, while the speed and interoperability roadmap is published by the Fibre Channel Industry Association. Four rules out of those documents decide most storage optics purchases.
1. Two generations back, not three
The FCIA roadmap states the rule in one line: each speed “maintains backward compatibility at least two previous generations”, and gives the worked example that 32GFC is backward compatible to 16GFC and 8GFC.
Read forwards that is reassuring. Read backwards it is a trap. A 32 Gigabit optic will not negotiate down to 4GFC, so a 32G module facing a 4G edge port leaves that port dark; a 16G module will, because 4G is two generations below it. Our catalogue reflects this exactly — the 16G designs below are specified for 4G, 8G and 16G, and the 32G designs for 16G and 32G only. A mixed-generation array refresh needs a port-by-port plan rather than a single part number. Fibre Channel transceivers covers the detail.
2. SW and LW, and a 32G part is an SFP28
Fibre Channel names its optics SW — short wave, 850 nm, multimode — and LW — long wave, 1310 nm, single-mode — rather than SR and LR. More consequentially, the FCIA speedmap puts 16GFC at 14.025 GBaud NRZ with throughput quoted as 3,200 MB/s, and 32GFC at 28.05 GBaud NRZ and 6,400 MB/s. Twenty-eight gigabaud is an SFP28 electrical interface, not SFP+.
So the thing that decides whether a 32G optic can work is the host’s electrical capability, not whether the module physically seats. An SFP+ port will take a 32G module mechanically and will not run it at 32G. On a storage platform that distinction is usually settled by the adapter generation rather than the chassis, which is one more reason to quote the host model with the part number. 16G and 32G Fibre Channel SFP+ works through it, and SFP+ vs SFP28 covers the Ethernet equivalent.
3. Multimode reach goes down as the generation goes up
This is the one that catches people during a refresh rather than during a design. The FCIA figures put 16GFC at 100 m on OM3 and 125 m on OM4, and 32GFC at 70 m on OM3 and 100 m on OM4, with OM4 required for 100 metres at 32G. Our own product records agree precisely, which means the figures are confirmed from two independent directions.
So an OM3 plant that was comfortable at 16G can fall short after a storage refresh, and the cable behind the patch panel looks identical either way. Single-mode has stayed at 10 km at every generation up to 32GFC, so where the span is long or the fibre grade unknown, long-wave removes the question entirely. Single-mode vs multimode fibre is worth reading before you commit, and why optical link budget matters covers the method where the reach is borderline.
4. At 32G an error counter means something different
32GFC introduced forward error correction to Fibre Channel, using a Reed-Solomon code over data transcoded from 64b/66b to 256b/257b. That is invisible in normal operation and decisive when a link is marginal: at 32G a healthy port is already correcting errors, so a non-zero corrected count is health rather than illness, where at 16G it would have been a warning. Forward error correction explained covers why, and spotting a failing fibre link before it drops covers what to watch instead.
Three more things that catch people out
1. Intel’s optics are adapter-side, and the X suffix is not documented in our records
The Intel parts in this range — E10GSFPSR, E10GSFPLR, E10GSFPSRX, E25GSFP28SR, E25GSFP28SRX, E25GSFP28LRX, E100GQSFP28SRX et E100GQSFP28LRX — are ordered against a network adapter, not a switch and not an array. The naming is the best in this guide: E, then the speed, then the cage, then the port type, with nothing to decode.
Two things about them need saying. First, the trailing X marks a variant and our records do not currently separate it: E10GSFPSR et E10GSFPSRX map to the same Carritech design with the same description, as do E25GSFP28SR et E25GSFP28SRX. Optically they are the same module. If the variant matters for your adapter’s support list, tell us which string the adapter expects and we will code to it. Second, an adapter’s support list is a different document from a switch’s, and a module that satisfies one does not automatically satisfy the other — which matters because the two ends of a server link are exactly that pair. Tell us both ends.
One Intel part breaks its own scheme entirely: SPTSHP2PMCDF is a 400 Gigabit QSFP-DD DR4 design with a twelve-character string that encodes nothing a reader can use. It is the exception that proves how good the rest of the family is.
2. Lenovo’s 800 Gigabit parts, and a DR8-versus-2×DR4 question
The newest parts in this range are Lenovo’s: 4TC7A83365, 4TC7B05746, 4TC7B09057 et 4TC7B10008, all 800 Gigabit OSFP modules for AI and high-performance computing hardware. Three optical designs between four order codes — 4TC7B05746 et 4TC7B10008 are the same 500 metre single-mode DR4 design under two numbers, which is the whole coding argument in miniature.
Two cautions. Our record for the SR8 part describes it as a QSFP-DD module while filing it as OSFP. It is an OSFP part, the cages are not interchangeable, and we are correcting the description — confirm the cage with us. And at 800 Gigabit there is a design question the part number does not answer: a DR8 part is one 800 Gigabit interface, while a 2×DR4 part is two independent 400 Gigabit interfaces in one cage, feeding two downstream devices from one faceplate position. Which you want is a fabric design decision rather than a speed upgrade, and the two are not substitutes. Tell us whether the port is configured as one 800G interface or two 400G ones. 800 Gigabit Ethernet port types come from IEEE Std 802.3df-2024, approved on 16 February 2024. Breakout explained covers the mechanics, QSFP-DD vs OSFP the form-factor choice, and AI and the new demand for high-speed connectivity covers why a storage vendor is now selling 800 Gigabit optics at all.
3. Sun and Oracle codings: we hold them, and you will have to ask
In the interests of being straight with you: Carritech holds eight Oracle-coded 16 Gigabit Fibre Channel optics, and this article does not link to them, because all eight product records are currently unreachable on this site. They are stocked and they are tested; the permalinks are broken, which is a problem on our side rather than a gap in the range. We are fixing the records.
In the meantime, if you are buying for a Sun or Oracle storage platform, send us the part number and we will quote it — the compatibility checker will find it even though the catalogue page will not. We would rather tell you that than quietly leave a vendor out of a guide that claims to cover the range.
Which cage takes which speed
| Cage | Speeds in this range | Where you find it |
|---|---|---|
| SFP+ | 1G, 10G, 16G FC | Array front ends, host bus adapters, SAN edge, server LOM |
| SFP28 | 10G, 25G, 32G FC | Current adapters, current array front ends |
| QSFP28 | 100G, or 4 × 25G broken out | High-rate adapters and converged fabrics |
| QSFP-DD | 400G, and QSFP28 optics in the same cage | Current high-end adapters |
| OSFP | 800G | AI and HPC hardware. Will not take a QSFP module of any kind |
Two notes for the order rather than the reading. A QSFP-DD cage accepts a QSFP28 module and an OSFP cage does not accept either, so the mechanical backwards compatibility runs one way and stops at OSFP. And a 32 Gigabit Fibre Channel module needs an SFP28 host port to run at 32 Gigabit, however comfortably it seats in an SFP+ one — which is the single most expensive misunderstanding in a storage optics order.
What the array checks, and what it does not
What an array, a host bus adapter or a SAN switch reads when a module seats is the identification data in the module’s EEPROM, laid out to the SFF management specifications held in the SNIA SFF specifications library, and compares it against what its firmware expects for that port. If the strings line up, the port comes up and the optic appears in the inventory like any other. If they do not, behaviour ranges from a log entry to a port that stays down.
On a storage fabric, a port that stays down is louder than a dark Ethernet uplink. It is a path that has gone missing from a multipath configuration, which gets noticed by monitoring, by the array, and frequently by an application team, all within minutes. That is the real reason storage optics are worth getting right first time rather than iterating.
What none of that hardware does is measure the optic. It cannot tell you whether the launch power is in specification, whether the receiver sensitivity matches the data sheet, or whether the part will still be in specification in three years. That is why every Carritech module is measured against its specification before despatch — see the 22 checks a module passes — and why coding and testing are two separate jobs. Fibre optic connector cleaning is the cheapest decibel you will ever recover in a storage room, where the patching is dense and the connectors get handled.
On the procurement questions, compatible optical transceivers vs OEM, OEM vs compatible: cost, risk and performance compared et do third-party optical transceivers void your warranty? answer them straight — and they are asked more often about storage than about anything else, because a support contract on an array is a bigger line than the array.
Five checks before you order
- Say what the optic plugs into. An array front end, a host bus adapter, a SAN switch and a server NIC have four different support lists, and the part number does not tell you which one issued it.
- Quote every identifier you can see — the feature code, the eight-character number, the string on the module label, and the machine type of the host. They describe the same optic in different languages and nobody maintains a translation table. We do, for the parts we stock.
- For Fibre Channel, count the generations. Two back works, three does not. A 32G optic will not serve a 4G port.
- Confirm the fibre grade before trusting a 32G multimode reach. OM4 for 100 metres; OM3 gives 70.
- Confirm the host’s electrical rate, not just the cage. A 32G module seats in an SFP+ port and will not run at 32G there.
The server and storage platform optics Carritech stocks
Fifty-three tested part numbers across Lenovo, IBM, NetApp and Intel codings, every one coded for the platform it is going into and measured against its specification before despatch. Listed by speed and form factor below, Fibre Channel first because that is what this range mostly is.
32G Fibre Channel SFP+ — 21 parts, the largest group
01KN789, 01KN793, 01KN795, 01KN799, 01YM942, 02CL041, 02PX560, 4M27A65416, 4M27A65417, 4M27A65418, 4M27A65419, 78P6616, 8960-2301, AJHE, AJHF, AJHJ, AJHK, X-000212, X-000213, X-000238, X-000239. All four naming schemes in one list: IBM’s feature codes and eight-character numbers, Lenovo’s ThinkSystem blocks, and NetApp’s rebadged Brocade numbers. These are SFP28 modules supporting 32G and 16G, short-wave at 850 nm over 100 m on OM4 and long-wave at 1310 nm over 10 km. Read the two-generation rule above before putting one of these into a mixed-generation fabric. Browse the 32G Fibre Channel range and the Émetteurs-récepteurs 32G page.
16G Fibre Channel SFP+ — 19 parts
00MY768, 00RY190, 00RY191, 00WC089, 03GH236, 332-00331, 332-00351, 4M27A65411, 4M27A65412, 4M27A65413, 4M27A65414, 88Y6393, 98Y2176, ACHU, X-48895-00-R6, X65403, X6596-R6, X6601A-R6, X6602A-R6. Supporting 16G, 8G and 4G — which makes these the right answer, not the old answer, wherever a fabric still has 4 Gigabit edge ports in it. Short-wave at 100 m on OM3 or 125 m on OM4, long-wave at 10 km. Note the two NetApp numbering generations side by side here: the 332-00nnn series and the X-prefixed one, with an -R6 revision suffix on several. Record the revision suffix as well as the stem — X6596-R6 et X6602A-R6 are different parts, not two spellings of one. Browse the 16G Fibre Channel range and the Émetteurs-récepteurs 16G page, or the combined 16G, 25G and 32G overview.
10G SFP+, 25G SFP28 and 100G QSFP28 — 8 parts, all Intel
10G duplex SFP+, 3 parts. E10GSFPLR, E10GSFPSR, E10GSFPSRX — SR on multimode and LR at 10 km. The port types come from IEEE Std 802.3ae-2002, approved at the June 2002 Standards Board meeting, and there is no forward error correction at 10G over fibre, which means an error counter on a 10G server link means exactly what it appears to mean — a useful property when you are trying to decide whether a storage performance problem is the network or the array. See SFP-10G-SR: the full specification and browse the 10G duplex SFP+ range and the 10G transceivers page.
25G SFP28, 3 parts. E25GSFP28LRX, E25GSFP28SR, E25GSFP28SRX — SR on multimode and LR at 10 km, from IEEE Std 802.3by-2016, approved on 30 June 2016. FEC configuration is the usual reason a 25G link stays down, it is a host setting rather than a module property, and on a server link that means agreeing it between the adapter driver and the switch — two different teams, frequently. Settle it before the maintenance window. See 25G SFP28 transceivers et forward error correction explained, and browse the 25G SFP28 range and the Émetteurs-récepteurs 25G page.
100G QSFP28, 2 parts. E100GQSFP28LRX, E100GQSFP28SRX — SR4 on multimode from IEEE Std 802.3bm-2015, approved on 16 February 2015, and LR4 at 10 km. At 100G and above forward error correction is part of the interface and permanently on, so a healthy port shows a steadily climbing corrected-codeword count rather than zero. These ports also break out to four 25G links where the platform supports the split, which on a server is frequently how a single adapter feeds four converged interfaces. See 100G QSFP28 transceivers and browse the 100G QSFP28 range and the 100G transceivers page.
400G QSFP-DD and 800G OSFP — 5 parts
400G QSFP-DD, 1 part. SPTSHP2PMCDF — a DR4 design reaching 500 m over single-mode. 400 Gigabit Ethernet was standardised in IEEE Std 802.3bs-2017, approved on 6 December 2017, with the faster electrical interfaces current parts use coming later from IEEE Std 802.3ck-2022, approved on 21 September 2022. Remember that a QSFP-DD cage will also take your existing QSFP28 optics, which makes a staged migration possible on an adapter you have already bought. See choosing 400G optics et the real TCO of 400G, and browse the 400G QSFP-DD range and the 400G transceivers page.
800G OSFP, 4 parts. 4TC7A83365, 4TC7B05746, 4TC7B09057, 4TC7B10008 — the Lenovo ThinkSystem range discussed above: a 2×VR4 design at 50 m on multimode, two order codes for one 500 m single-mode DR4 design, and an SR8 design at 50 m. Read the DR8-versus-2×DR4 caution above before ordering. One commercial point worth making here, because this is the part of the range where it bites hardest: order 800 Gigabit spares with the production modules rather than after the first failure. The supply base at the top of the range is narrower than it is at the bottom, and an AI or HPC cluster waiting on a module is an expensive thing to have idle. See 800G optical transceivers, 800G OSFP optical transceivers, 800G and 1.6T migration et optical transceiver power consumption, and browse the 800G OSFP range and the 800G transceivers page. Our guide to NVIDIA compatible optics covers the rest of the AI fabric range, including the lid geometry question that decides whether an 800G OSFP module physically seats.
Inside the rack, do not buy optics at all
This is the range where that advice applies most and is taken least. Server and storage estates are rack-dense by definition: the adapter and the top-of-rack switch are two feet apart, and the array and the fabric switch are frequently in the same cabinet. For those links two transceivers and a patch lead cost more than a cable, draw power at both ends, add two connectors to keep clean and two optical budgets to manage.
Les 10G SFP+, 25G SFP28, 100G QSFP28 et 800G OSFP direct attach ranges cover exactly that case, with active optical cable where the run is longer or the copper too stiff for the cable management behind a dense rack. See DAC and AOC cables et DAC vs AOC.
One point specific to this range: a cable assembly has an EEPROM at each end, and each end faces a different vendor. A Lenovo or Intel adapter at one end and a Cisco, Arista or Brocade switch at the other means the assembly needs correct identification twice, to two different firmware families’ expectations — and the server end and the switch end are almost never the same vendor. Plenty of suppliers can code one end. Coding both is what the Carritech Opticode coding box does at the point of deployment, and it is the difference between a cable that works and a cable that comes up on one side only.
Getting storage platform optics without the OEM price
Carritech Optics holds the Lenovo, IBM, NetApp and Intel coded range in UK and EU stock, from 10G et 25G through 100G et 400G to 800G, with the 16G et 32G Fibre Channel ranges that make up three-quarters of it. Patch cords, copper SFPs and adapters et media converters are alongside.
Every module is coded for the target platform and measured against its specification before despatch, and carries a garantie à vie with UK and EU stock and support behind it. The stock argument has a particular shape in a storage estate: there are only four optical designs in the whole Fibre Channel part of this range — short-wave and long-wave at each of two generations — sitting behind forty part numbers from four vendors. Holding one optical design per reach and coding at the point of deployment collapses forty line items into four, and it is the clearest illustration of the case made in building a smarter transceiver stock strategy. Your optics will go end-of-sale before your switch does applies doubly to storage, where arrays are bought on five-year support contracts and kept for eight. Our accreditations, why choose us et support pages cover how we work, and the knowledge centre holds the rest of the technical library. Our guides to Brocade, Dell et HPE and Aruba compatible optics cover the other vendors in a typical storage estate, and compatible optics for Cisco platforms covers the fabric side.
Trying to match an IBM feature code to something orderable, working out whether a NetApp X- number is a Brocade part, or establishing whether a 32G optic will serve a 4G edge port? Tell us the feature code or part number, the machine type and what is at the far end and we will say. Our compatibility checker returns the tested Carritech equivalent of an OEM part number in seconds, upload a whole part-number list and we will price it, or request a quote.