Skip to content

Technology card

Shipping

PCIe Storage

PCI Express storage

Separate the interconnect from NVMe, calculate coding-adjusted link ceilings, and diagnose negotiated speed, lane width and shared upstream paths.

Glossary entry on hesela.dev · JSON record · PCIe lanes · Bifurcation · PCIe switch · Retimer · Redriver

On this page6 sections

In one minute

PCIe storage connects a storage controller to a host through PCI Express links, with NVMe commonly providing the storage interface independently of the drive's physical form factor.

Signal conditioning solves a different problem from adding lanes. A redriver uses analog equalization and gain to compensate channel attenuation. It does not recover the clock, so it cannot reset incoming random jitter. A retimer recovers data and timing before retransmission and participates in physical-layer training. Neither mechanism is a bandwidth multiplier. [1]

Linear and limiting redrivers also differ: linear behavior preserves amplitude relationships within the operating range, whereas limiting output is nonlinear. Added de-emphasis in a limiting design can interfere with link training. Use the actual component model and channel measurements when choosing or tuning a device; the label alone does not establish suitability. [2]

A retimer does not replace a PCIe switch. Its physical-layer segments lack the Data Link and Transaction Layers used for packet routing. Evaluate loss and compliance separately on each side; do not translate two segments into a guaranteed doubling of cable or trace length. The cited PCIe 5.0 description requires matching widths on the two sides, even when configured narrower. System-level interoperability still needs verification. [3]

Equalization bypass does not disable a retimer: the cited PCIe 5.0 Q&A says it still signals its presence during training. Conversely, disabling an active retimer or redriver stops data passage; this is not a passive bypass path. Check the documented control before interpreting a link-loss test. [3]

Measurement endpoints matter. Neugebauer and colleagues' SIGCOMM 2018 pcie-bench study used programmable NIC/FPGA implementations on Gen3 x8 hosts, not SSD workloads. Because PCIe Memory Writes are posted, its LAT_WRRD probe times a write followed by a read; it does not isolate write latency. Report this compound observation as such, not as an SSD write or persistence latency. [4]

A switch also does not establish that an application can perform peer-to-peer DMA. Linux v6.12 documents separate resource-provider, client and orchestrator roles, plus a topology compatibility check. Routing between different root-port hierarchy domains is not generally required; the documented compatibility rule includes same-root-port devices or specifically allowlisted host bridges. Check the actual driver and platform path rather than infer support from lane width. [5]

Support for one narrower x4 device does not imply support for four independent x4 links. Intel's version-009 processor datasheet distinguishes narrower-device support from further bifurcation. Check the CPU controller, selected slot routing and firmware configuration together; the example does not qualify other processor families. [6] [7]

PCIe, NVMe and the enclosure answer different questions. PCIe is the interconnect; NVMe defines how host software communicates with non-volatile storage and includes separate transport specifications. M.2, U.2 and add-in cards describe physical implementations, not equivalent protocol names. NVMe can also use RDMA or TCP, so NVMe does not imply a local PCIe connection. [8]

PCIe storage does not necessarily mean NVMe. Samsung's 2015 SM951 announcement explicitly distinguishes its new NVMe model from the earlier AHCI-based PCIe version. This is a historical counterexample to treating the two names as synonyms, not a recommendation for either historical device. [13]

Describe an operating link with both rate and width. Gen3, Gen4 and Gen5 correspond to 8, 16 and 32 GT/s per lane. A negotiated x4 link uses four lanes. For these 128b/130b links, the table computes rate times four times 128/130 divided by eight, in decimal GB/s for one direction. The factor of eight converts bits to bytes; GB means 1,000,000,000 bytes. [9] [10]

This calculation removes only the coding overhead. It does not subtract all packet, link-management or storage-protocol costs, and it is not an SSD benchmark. Do not double the result for a one-direction read figure. The table is deliberately limited to Gen3/4/5; do not reuse its encoding factor for PCIe 6.0 or 7.0 FLIT operation. [9] [10]

Inspect the negotiated state rather than only the product label. Linux v6.12 implements current_link_speed and current_link_width by reading the PCIe Link Status register, with separate maximum-capability attributes. For the correct PCI device address, these read-only attributes distinguish the observed link from its advertised limits; they do not measure completed storage I/O. Record the observation time and operating conditions. [12]

Follow the path beyond the SSD. Linux's pcie_bandwidth_available walks upstream to find the minimum-bandwidth link. That identifies a link-configuration ceiling along the traversed upstream PCIe path, not unused bandwidth or measured SSD throughput. Intel separately explains that chipset-connected devices share the CPU-chipset connection. A fast endpoint can therefore coexist with a constrained or busy upstream path. [11] [16]

Bifurcation divides a host connection into multiple independent links for a supported wiring arrangement. StarTech's DUAL-M2-PCIE-CARD-B requires host support for this operation; its PEX8M2E2 instead uses an ASM2824 switch and does not require host bifurcation. These named products demonstrate the distinction, not a universal compatibility promise for adapters, motherboards or firmware. [15] [14]

For a performance investigation, record the SSD and firmware, host slot, adapter, negotiated speed and width, upstream topology, and whether other devices were active. Then report workload, request size, queue depth and measured throughput separately from the calculated interface ceiling. This is an evidence checklist, not a claim that link width alone predicts latency or sustained flash performance.

Link-rate ceilings

Derived values for a negotiated x4 link, not measured payload throughput. Formula: GT/s per lane x 4 x 128/130 / 8. Remaining protocol overhead is not deducted. [9] [10]

Gen3/4/5 x4: coding-adjusted ceilings per direction, before remaining protocol overhead
Link generationTransfer rate per lanex4 ceiling (decimal GB/s)
PCIe 3.08 GT/s3.938
PCIe 4.016 GT/s7.877
PCIe 5.032 GT/s15.754

Uncertainty notes

  • Source-based engineering explanation with reproducible arithmetic, not a hardware qualification test or SSD performance measurement.
  • The numeric table covers only conventional Gen3/4/5 128b/130b links. No current-generation leadership, throughput guarantee or universal platform compatibility is inferred.
  • Linux observations are grounded in v6.12 source. Vendor examples establish only the named products' documented designs.

Sources

  1. 01PCI-SIG, September 2019: PCI Express retimers versus redrivers · accessed 2026-11-19
  2. 02Texas Instruments SNLA401, December 2021: PCIe signal conditioning and linear versus limiting redrivers · accessed 2026-11-19
  3. 03PCI-SIG, January 2021: PCIe 5.0 implementation Q&A, questions 1, 3, 5, 7, 8 and 26 · accessed 2026-11-20
  4. 04Neugebauer et al., SIGCOMM 2018, Understanding PCIe performance for end host networking, sections 4-6; DOI 10.1145/3230543.3230560. NIC/FPGA measurements, not an SSD benchmark. · accessed 2026-11-03
  5. 05Linux v6.12: PCI Peer-to-Peer DMA Support; hierarchy, compatible providers and driver roles. · accessed 2026-11-03
  6. 06Intel 13th/14th Gen Core datasheet v009: narrower-link support versus lane bifurcation · accessed 2026-11-02
  7. 07ASUS Hyper M.2 compatibility: CPU, slot selection and firmware configuration · accessed 2026-11-02
  8. 08NVM Express: specification set overview; command sets and transports are separate from form factors. · accessed 2026-10-18
  9. 09PCI-SIG: PCIe 7.0 educational webinar, July 2025, slide 7; historical Gen3/4/5 rates and encoding. No later-generation throughput is calculated here. · accessed 2026-10-18
  10. 10PCI-SIG: How does PCIe 3.0 8 GT/s double PCIe 2.0 5 GT/s? Encoding efficiency and direction conventions. · accessed 2026-10-18
  11. 11Linux PCI support library: pcie_bandwidth_available; upstream path and limiting link. · accessed 2026-10-18
  12. 12Linux v6.12, drivers/pci/pci-sysfs.c: current_link_speed_show and current_link_width_show read Link Status; maximum attributes read capabilities. · accessed 2026-10-18
  13. 13Samsung, 15 April 2015: SM951 NVMe announcement identifies an earlier AHCI-based PCIe version. Historical counterexample, not current product advice. · accessed 2026-10-18
  14. 14StarTech PEX8M2E2: ASM2824 switched dual-SSD adapter; no host bifurcation required for this product. · accessed 2026-10-18
  15. 15StarTech DUAL-M2-PCIE-CARD-B: dual-SSD adapter explicitly requiring host PCIe bifurcation; product-specific compatibility requirements. · accessed 2026-10-18
  16. 16Intel: What is PCIe 4.0 and why does it matter? Chipset-attached devices share the CPU-chipset path; example link specifications are platform-specific. · accessed 2026-10-18

Published by Hesela team

Report a correction · Editorial standards · Reuse guidance

Cite this page
Hesela team. PCIe Storage. Hesela. https://hesela.com/tech/pcie-storage/