SlimSAS SFF-8654: Bandwidth, Use Cases and NVMe Compatibility

Illustration — Indrajit Das / Wikimedia Commons (CC BY-SA 3.0)

SlimSAS (SFF-8654) is the high-density cable and connector family that carries PCI Express signals inside modern servers and storage — the physical layer underneath U.2, U.3, OCuLink and most Gen4/Gen5 NVMe backplanes. If you are speccing a storage chassis, the connector you choose sets the bandwidth ceiling and the signal-integrity budget, so it is worth understanding exactly what SlimSAS does and does not do before you order a backplane or a cable. This guide goes past the marketing and into the numbers that actually decide whether your NVMe links stay at full width.

Key takeaways

• SlimSAS (SFF-8654) carries PCIe (and tunneled SAS/SATA) at up to 8 lanes in a tiny footprint — it is the connector behind most modern NVMe backplanes.
• Per-lane bandwidth: PCIe Gen4 = 16 GT/s, Gen5 = 32 GT/s; an 8i cable therefore moves ~16 GB/s at Gen4 and ~32 GB/s at Gen5.
• It is not the same as Mini-SAS HD (SAS-focused) or OCuLink (SFF-8611, external); match the connector to your backplane's pinout and generation.
• Buy on signal-integrity rating and certified cable length, not price — a marginal cable is the most common cause of NVMe link drops.

What SlimSAS actually carries

SlimSAS was defined by the SFF Committee in specification INF-8077 (SFF-8654). It is a small-form-factor, high-speed connector system designed for PCIe signalling at up to 8 lanes, with versions that also tunnel SAS and SATA through PCIe-SAS bridge logic. In practice it is the cable you plug between a motherboard or RAID/HBA and an NVMe drive bay.

Because it is a pure transport for the PCIe protocol, the 'SlimSAS' name says nothing about the drive protocol on the other end — that is decided by the backplane. A U.2 / U.3 (SFF-8639) bay, for example, receives SlimSAS signalling and presents NVMe to the OS, while a SAS backplane using the same physical cabling relies on bridge logic to translate.

• PCIe Gen4 / Gen5 lanes for NVMe SSDs (x4 per drive is the dominant pattern).
• SAS / SATA tunneling where a backplane mixes protocols behind a bridge.
• Up to 4 lanes on the common 4i variant and up to 8 lanes on 8i (4i = 38 positions, 8i = 74 positions).
• 85 ohm differential controlled-impedance signalling with tight return-loss limits.

Variant	Lanes	Typical use
SFF-8654 4i	x4	Single NVMe drive link, compute-to-drive
SFF-8654 8i	x8	Dual-drive or x8 compute fabric
SFF-8654 4i+4i	2 x4	Two independent NVMe links in one connector
SFF-8654 8i+8i	2 x8	High-density compute/storage fabric

Bandwidth: what the generations really deliver

Bandwidth is set by the PCIe generation, not by the connector family. PCI-SIG defines the raw lane rate; after 128b/130b encoding the usable throughput is slightly lower. The connector and cable only matter insofar as they preserve signal integrity at that rate — a compliant Gen5 channel is allowed far less loss than a Gen4 one.

A common mistake is to assume any cable stamped 'SlimSAS' performs the same. At Gen5 the eye diagram closes quickly with cheap dielectric or poor shielding, so cable certification and length are part of the spec, not an afterthought. The numbers below are per-direction, per-lane, and assume a fully compliant channel.

• PCIe Gen4: 16 GT/s per lane (~16 Gbps raw, ~15.75 Gbps encoded).
• PCIe Gen5: 32 GT/s per lane (~32 Gbps raw, ~31.5 Gbps encoded).
• 4i (x4) Gen4 ~ 64 Gbps; 8i (x8) Gen4 ~ 128 Gbps (~16 GB/s).
• 8i (x8) Gen5 ~ 256 Gbps (~32 GB/s) over a single connector.

Configuration	Gen4 throughput	Gen5 throughput
x4 (4i)	~64 Gbps / ~8 GB/s	~128 Gbps / ~16 GB/s
x8 (8i)	~128 Gbps / ~16 GB/s	~256 Gbps / ~32 GB/s
x16 (via MCIO/SFF-TA-1016)	~256 Gbps / ~32 GB/s	~512 Gbps / ~64 GB/s

Signal integrity and the cable-length budget

At 16 GT/s and especially 32 GT/s, the cable is a transmission line, not a wire. Insertion loss, return loss and crosstalk between adjacent pairs determine whether the receiver recovers the signal. Vendors express this as an insertion-loss budget in decibels at the Nyquist frequency; a Gen5 cable must stay within a much tighter budget than a Gen4 one, which is why the same physical connector supports different maximum lengths by generation.

In practice this means you should treat the cable as a signal-integrity component and ask for the measured S-parameters, not a generic 'supports Gen5' claim. A cable that is fine at 0.5 m may fail compliance at 1 m, and the failure mode is insidious: the link trains at x4 instead of x8, or drops width under load, rather than refusing to come up at all.

• Request insertion-loss / return-loss plots at the target generation.
• Keep runs inside the vendor-certified length for that generation.
• Route away from high-frequency noise sources; maintain pair twist.
• Budget margin: do not design to the exact limit you measured.

Where SlimSAS is the right choice

Use SlimSAS when the link is PCIe-native and density matters — high-port-count NVMe backplanes, GPU/compute fabric cabling, and enterprise storage where you need many lanes in limited space. Its low profile is exactly why it displaced larger legacy connectors in 1U/2U systems, where every millimetre of connector height is contested by airflow and neighbouring cards.

It pairs naturally with OCuLink (SFF-8611/8612) for external PCIe, and with MCIO (SFF-TA-1016) when you step up to Gen5/Gen6 and x16 widths. Many modern servers therefore carry a mix: SlimSAS internally, OCuLink or MCIO at the edge.

• High-density NVMe backplanes (U.2 / U.3).
• Compute-to-accelerator and storage fabric inside 1U/2U nodes.
• Anywhere signal integrity at high lane count is non-negotiable.

SlimSAS vs the alternatives

SlimSAS is frequently confused with Mini-SAS HD and OCuLink. They overlap in physical size but serve different protocols and placements. Picking the wrong one means a cable that physically will not seat or, worse, seats but violates the pinout and risks damage.

Connector	Primary protocol	Placement	Note
SlimSAS SFF-8654	PCIe (NVMe)	Internal	High-density, Gen4/Gen5
Mini-SAS HD SFF-8643/8644	SAS / SATA	Internal / external	SAS backplanes, 12/24 Gbps
OCuLink SFF-8611/8612	PCIe	External	External NVMe enclosures
MCIO SFF-TA-1016	PCIe	Internal	Gen5/Gen6, up to x16

Specifying it without surprises

Treat the cable as a signal-integrity component, not a wire. The single biggest cause of intermittent NVMe link-width drops is a non-compliant or over-length cable. Ask the vendor for the insertion-loss and return-loss plot at your target generation, and keep the run inside the certified length. For the connector's published lane counts, bandwidth and impedance figures, see the SlimSAS SFF-8654 technical guide.

• Confirm the exact 4i / 8i pinout your backplane expects.
• Match cable generation to board generation (a Gen4 cable on a Gen5 board loses margin).
• Respect the certified length — Gen5 budgets are tighter than Gen4.
• Verify shielding/ESD handling for your install environment.

Decision	Why it matters
Pinout (4i/8i)	Wrong pinout will not seat or will mis-wire lanes
Generation match	Mismatched cable steals Gen5 margin
Certified length	Over-length causes width drops under load
SI documentation	Plot proves compliance beyond a label

Gen4-to-Gen5 migration reality

When you move a backplane from Gen4 to Gen5, the connector often stays SFF-8654 but the cable and the channel budget change. The board, the retimer or redriver (if any), and the cable must all certify at 32 GT/s. Many 'Gen5-ready' chassis are shipped with Gen4 cables that quietly cap performance until the cable is swapped — a detail worth checking in any upgrade project.

If your roadmap includes Gen5, buy Gen5-certified cables from day one even if the first wave of drives is Gen4. The cable is the cheapest component to get right once, and the most expensive to discover wrong after a rack is cabled.

• Gen5 needs certified cable + compliant channel end to end.
• Retimers/redrivers must also certify at 32 GT/s.
• Buying Gen5 cable early avoids a later rip-and-replace.

Pinout and mating variants in practice

SlimSAS is not one connector but a family of keying and lane-count combinations. The two you will meet most often are 4i (a single x4 PCIe link) and 8i (a single x8 PCIe link), plus combo variants such as 4i+4i and 8i+8i that carry two independent links in one shell. Because 8i is a superset of two 4i links in many backplanes, an 8i host port is frequently broken out to two U.2 drives with a breakout cable — a pattern worth knowing before you assume a single drive per connector.

Keying matters. The physical latch and the key position differ between variants, and forcing the wrong plug risks damaged pins. Always match the connector on the host board to the connector on the backplane and the cable in between; the three must agree on both lane count and keying. When in doubt, ask the vendor for the exact SFF-8654 variant reference printed on the part.

• 4i = one x4 PCIe link; 8i = one x8 PCIe link.
• Combo variants (4i+4i, 8i+8i) carry two links in one shell.
• 8i host ports are often broken out to two U.2 drives.
• Keying and latch differ by variant — never force a plug.

Thermal and mechanical factors

A connector choice is also an airflow decision. SlimSAS was designed to be low-profile precisely so it does not obstruct the high-velocity air a 1U server relies on. Its height sits far below legacy connectors, which is one reason modern dense storage nodes can pack more drives without starving the CPU heatsink.

Mechanically, the cable needs strain relief and a sensible bend radius. Repeated sharp bends at the connector tail are a common cause of intermittent faults that look like drive failures but are actually cable damage. Route with gentle radii, secure the cable near the connector, and avoid hanging weight on the plug.

• Low profile preserves the airflow budget of 1U/2U nodes.
• Respect minimum bend radius at the connector tail.
• Add strain relief; never let the cable hang on the plug.
• Secure routing to avoid vibration fretting in mobile racks.

Interoperability testing and what to demand

Compliance is demonstrated, not asserted. A credible vendor tests the cable's S-parameters against the PCIe channel budget at the target generation and provides the insertion-loss and return-loss plots. Increasingly, vendors also run interoperability ('plugfest') testing with specific host and backplane combinations, since a cable that passes a bench test can still misbehave in a particular chassis.

For a deployment, ask for the compliance report at your exact generation and, where possible, references from a comparable chassis. The cost of one bad cable lot — in downtime and in the engineering hours to diagnose it — dwarfs the unit price difference between a certified and an uncertified cable.

• Demand S-parameter plots at the target PCIe generation.
• Prefer vendors with interop testing on real chassis.
• Request references from a similar deployment.
• Treat the cable as a spec'd component, not a commodity.

Specifications as listed on link-lungkay.com

Pulled directly from the manufacturer's published page (SlimSAS SFF-8654 Guide: High-Speed NVMe Cable Connectivity); we do not reconstruct or estimate these values.

Parameter	SlimSAS 4i (SFF-8654)	SlimSAS 8i (SFF-8654)
Lane Count	4 Lanes (x4)	8 Lanes (x8)
Bandwidth per Lane	24Gbps (SAS 4.0) / 16GT/s (PCIe 4.0)	24Gbps (SAS 4.0) / 16GT/s (PCIe 4.0)
Total Bandwidth	Up to 96Gbps	Up to 192Gbps
Impedance	85 Ω (PCIe) / 100 Ω (SAS)	85 Ω (PCIe) / 100 Ω (SAS)
Pin Count	38-pin	74-pin
Insertion Loss	< -5dB @ 8GHz	< -5dB @ 8GHz

Buyer's specification checklist

• Backplane pinout confirmed as 4i or 8i before ordering.
• Cable generation matches board generation (Gen4 vs Gen5).
• Certified cable length within signal-integrity budget.
• Vendor supplies insertion-loss / return-loss data at target gen.
• Connector shielding and ESD handling specified for the environment.

Frequently asked questions

How many lanes does a SlimSAS cable carry?

A 4i connector carries four PCIe lanes and an 8i connector carries eight, which is why lane count determines bandwidth rather than the connector's physical size.

Can I run NVMe over a longer cable?

Only within the length certified for the generation you are running. Exceeding it degrades signal integrity and shows up as intermittent link-width drops rather than a clean failure.

What causes a link to drop to x1?

Almost always a non-compliant or over-length cable, or a marginal connector mate. Check the insertion-loss plot and reseat the connector before blaming the drive.

What is the difference between SlimSAS and Mini-SAS HD?

Mini-SAS HD is the older, larger connector family; SlimSAS is smaller and aimed at higher-density PCIe runs. Density, not bandwidth alone, is usually why a design moves to SlimSAS.

Can I mix cable vendors on one backplane?

You can, but each cable must meet the same insertion-loss and impedance specification. Mixing marginal and compliant cables makes intermittent faults very hard to diagnose.

How do I specify a cable for PCIe Gen5?

By insertion loss at the target frequency, impedance tolerance and certified length for that generation. The generation name on its own is not a specification.

Illustration — User:ParkerHiggins (Electronic Frontier Foundation) / Wikimedia Commons (CC0)

References and standards

• SFF Committee, SFF-8654 (INF-8077) connector definition.
• PCI-SIG, PCI Express Base Specification — lane signaling rates.
• SNIA / T10 storage connector interoperability notes.

Standards and references. Serial Attached SCSI; SNIA; ISO 9001