Adapters and embedded modules
PCIe host, target and NTB products across server, workstation, XMC, PXIe and CompactPCI Serial environments.
Dolphin Interconnect Solutions
Dolphin PCIe Expansion and Low-Latency Fabric Solutions
Primionics configures Dolphin PCIe adapters, switches, expansion systems, cables and eXpressWare software for transparent external I/O, independent-host NTB communication, shared memory, reflective memory, peer-to-peer data paths and controlled access to remote accelerators.
PCIe 3.0 / 4.0 / 5.0
Transparent & NTB
Copper & optical links
Hardware + software
Two nodes to switched fabrics
One platform — multiple system models
Host / CPU node A
x86, ARM or embedded compute
Host / CPU node B
Independent root complex
PCIe adapter, NTB or switch fabric
Direct, fan-out, star, mesh, multi-switch or expansion topology
GPU / FPGA
NVMe / DAQ
Remote I/O
What Dolphin provides
From external I/O expansion to multi-host PCIe fabrics.
A Dolphin system can be a simple cable between a workstation and an external device, a deterministic shared-memory network between real-time computers, a switched multi-host fabric, or a composable pool of GPUs, FPGAs and NVMe resources.
System behaviour depends on the complete topology.
Reliable operation depends on transparent or NTB mode, topology, lane width, BAR and reset behaviour, clocking, cabling, operating system, IOMMU configuration, software API and failure recovery. These elements must be validated together.
Switches and expansion
External switches, active and passive backplanes, and eBox platforms for remote I/O and scalable fabrics.
Copper and optical cabling
SFF-8644, SFF-8614, CopprLink/CDFP and optical link options selected by generation, distance and topology.
eXpressWare and application APIs
Shared memory, sockets, IP, SmartIO, reflective memory, peer-to-peer transfers, diagnostics and management.
Configuration model
A reliable PCIe system begins with five architectural decisions.
Dolphin’s flexibility comes from combining each layer independently. A high-bandwidth card alone does not determine whether the system behaves as expansion I/O, a network, shared memory or a resource pool.
Hosts and processors
One or more servers, workstations, embedded CPUs, SoCs, PXIe controllers or rugged compute nodes.
Transparent or NTB
Choose local-style device enumeration or isolated host address spaces with managed communication.
Direct or switched
Use point-to-point, fan-out, mesh, star, expansion chassis or one/more external switches.
I/O and accelerators
Connect GPUs, FPGAs, NVMe, frame grabbers, DAQ cards, network adapters or custom PCIe devices.
Access model and APIs
Use standard drivers, SISCI, SuperSockets, IPoPCIe, SmartIO, reflective memory or an OEM stack.
Architecture choices
Select the PCIe model that matches ownership, isolation and data movement.
These configurations solve different problems. Treating them as interchangeable creates driver, enumeration, isolation or real-time issues later in the project.
Transparent PCIe expansion
A host accesses remote PCIe devices as though they were installed in its local chassis. Suitable for adding slots, relocating noisy or high-power devices, and placing I/O near the equipment under test.
Host-to-host NTB networking
Non-transparent bridging separates independent hosts while enabling high-throughput, low-latency communication between their memory domains. Suitable for tightly coupled processing and distributed embedded systems.
Switched PCIe fabrics
External switches connect multiple hosts and endpoints using star, fan-out or multi-switch designs. Port width and allocation can be adapted to node count, bandwidth and redundancy requirements.
Reflective memory and multicast
Updates from one node are distributed into mapped memory regions on multiple nodes without a traditional network protocol in the runtime data path. Used where deterministic distribution and low jitter matter.
SmartIO and Device Lending
PCIe resources such as GPUs, FPGAs and NVMe devices can be made available to other systems and reassigned without physically moving the card. This enables composable and better-utilised infrastructure.
Peer-to-peer device data paths
Compatible PCIe masters can move data directly between devices—such as FPGA to GPU or acquisition card to accelerator—reducing intermediate CPU copies and memory traffic.
Complete portfolio
A complete stack from host adapters to fabric software.
The portfolio spans several generations and form factors so existing platforms can be extended while new systems move to higher bandwidth and longer-distance links.
NTB host adapters
Adapters for host-to-host communication, shared memory, DMA, reflective memory and switched multi-computer networks.
- PCIe x4, x8 and x16 options
- Direct, mesh and switched use
- eXpressWare software support
Transparent host / target adapters
Host and target functions for external PCIe I/O, device relocation and expansion systems using standard PCIe enumeration.
- Remote devices appear locally
- Host clock isolation options
- Copper and optical variants
PCIe switches
Rack-mounted and embedded switch platforms for fan-out, star, scalable fabrics and mixed host/device configurations.
- Transparent and NTB firmware modes
- Port partitioning by application
- Single- and multi-switch fabrics
eBox expansion systems
External chassis for installing standard add-in cards outside the host, including high-power GPU, FPGA, storage and acquisition configurations.
- Multi-slot PCIe expansion
- Single or multiple uplinks
- Monitoring and management options
Backplanes and embedded form factors
Platforms for OEM, test and rugged systems where conventional server cards are not the correct mechanical format.
- PXIe modules and switching
- XMC and CompactPCI Serial
- VPX and custom/OEM support
Cables and optical links
External PCIe cabling selected by generation, lane width, connector, bend constraints, electrical reach and installation distance.
- SFF-8644 and SFF-8614
- CopprLink / CDFP
- Optical links for longer reach
Product-family reference
| Platform family | NTB / networking examples | Transparent / expansion examples | Typical role |
|---|---|---|---|
| PCIe 5.0 host adapters | MXH530, MXH570 | MXH532, MXH572 | High-bandwidth x16 cabled host links, direct networks, expansion and fabric uplinks. |
| PCIe 4.0 host adapters | MXH914, MXH916, MXH918, MXH930, MXH940, MXH950 | MXH915, MXH917, MXH919, MXH932, MXH942, MXH952 | x4/x8/x16 copper, fan-out and optical configurations for networking or remote I/O. |
| PCIe 3.0 host adapters | MXH910, MXH830; PXH810, PXH830, PXH840 | MXH912, MXH832; PXH812, PXH832, PXH842 | Established server, workstation and long-lifecycle PCIe 3.0 deployments. |
| XMC modules | PXH820 | PXH822 | Rugged and embedded PCIe networking or transparent I/O in XMC systems. |
| PXIe modules | MXP908 | MXP909; MXP924 switch module | PXIe chassis integration, peripheral connectivity and test-system fabrics. |
| CompactPCI Serial | MXC960 | MXC962, MXC948 | Embedded PCIe and Gigabit Ethernet switching in CompactPCI Serial platforms. |
| External PCIe switches | MXS524, MXS924, MXS824 | Scalable transparent or NTB fabrics, fan-out, clustering and reflective-memory networks. | |
| Expansion platforms | eBox 4, eBox 4 Pro, active and passive PCIe backplanes | Remote multi-slot I/O, accelerator expansion, multi-uplink and rack-mounted systems. | |
The table is a family-level guide, not a substitute for configuration validation. Exact availability, topology, cable length, lane allocation, operating system and software features must be confirmed for the selected hardware revision.
eXpressWare software
Choose the software model that matches latency, compatibility and control.
Applications can use familiar sockets and IP, direct memory mapping, DMA, multicast, device access or management tools. The correct layer depends on how much control and determinism the application needs.
Applications
Simulation, imaging, acquisition, analytics, storage, control, clustering and OEM software.
SuperSockets
Berkeley-sockets-compatible acceleration for applications that need low latency without a complete rewrite.
IPoPCIe
Standard TCP/IP networking over the PCIe interconnect for conventional network software and services.
SISCI API
Shared/remote memory, DMA, PIO, interrupts, reflective memory, multicast and direct device access.
SmartIO
Remote PCIe device access, hot add, device lending and resource allocation for compatible deployments.
IRM / GENIF
Interconnect resource management, mappings, interrupts, DMA, node events, heartbeat and fabric bring-up.
Hardware layer
Dolphin adapters, switches, expansion systems, embedded modules and supported OEM PCIe switch platforms.
Deployment topologies
Scale from one remote slot to a multi-node fabric.
The same technology family can support very different physical and logical layouts. Bandwidth per node, failure isolation and software complexity change with each topology.
Transparent point-to-point
One host connects to one remote device, target adapter or expansion chassis.
Two-node NTB
Independent hosts communicate through shared memory, DMA, sockets or IP.
Switched star / fan-out
Multiple hosts or endpoints connect through a centrally managed PCIe switch.
Multi-switch fabric
Larger node counts, segmented bandwidth and complex OEM or laboratory systems.
Industrial and scientific use cases
Low-latency data movement for test, embedded and accelerated computing.
The use case is not merely “faster data.” The value is often lower latency, deterministic distribution, remote placement, reduced copies, resource sharing or cleaner separation between compute nodes.
High-rate data acquisition
Place digitizers, frame grabbers and FPGA I/O near the test object while processing data in a separate host or accelerator system.
HIL and real-time systems
Use NTB shared memory or reflective memory to exchange state data between simulators, controllers and I/O nodes with low jitter.
Machine vision and analytics
Move image or sensor streams directly from acquisition hardware to GPU or FPGA processing without unnecessary CPU copies.
Distributed mission computing
Interconnect rugged and embedded processors, sensors and accelerators using XMC, VPX or cabled PCIe architectures.
Imaging and diagnostics
Connect acquisition, reconstruction and display compute nodes where predictable high-throughput data movement is critical.
GPU and FPGA infrastructure
Expand accelerator capacity, establish peer-to-peer paths or create pools of devices that can be assigned to different hosts.
NVMe over PCIe fabrics
Build high-performance local fabrics and storage pools where standard network protocol overhead is undesirable.
ADAS and central compute
Apply static PCIe communication, diagnostics and controlled resource mapping using the eXpressDrive architecture for automotive programmes.
Distributed video processing
Transport high-rate uncompressed or processed video between capture, compute, storage and output systems.
Instrument and compute clusters
Couple acquisition instruments, memory and compute nodes with low-latency host-to-host and device-to-device communication.
Ultra-low-latency data paths
Accelerate tightly controlled communication and failover paths where every microsecond affects transaction performance.
Custom PCIe products
License software or adapt switch-based designs for proprietary boards, embedded systems and long-lifecycle products.
Selection guide
Select the architecture from system behaviour and compatibility requirements.
This matrix narrows the architecture. Final selection still requires lane budgeting, PCIe hierarchy review, endpoint compatibility and software validation.
| Requirement | Likely architecture | Core hardware | Software model | Critical checks |
|---|---|---|---|---|
| Add PCIe slots outside a workstation | Transparent expansion | Host adapter + target/uplink + backplane or eBox | Standard endpoint driver; optional board management | Enumeration, power, cooling, reset behaviour, cable reach |
| Exchange data between two independent computers | Direct NTB connection | NTB adapter in each host | SISCI, SuperSockets or IPoPCIe | OS support, address mapping, message size, latency target |
| Connect several computers with low latency | Switched NTB fabric | NTB adapters + external PCIe switch | eXpressWare + network management | Port allocation, oversubscription, failover, node count |
| Distribute the same state to many real-time nodes | Reflective memory / multicast | NTB adapters + multicast-capable switch | SISCI reflective memory | Update size, frequency, jitter, memory consistency, topology |
| Move acquired data directly to a GPU or FPGA | Peer-to-peer PCIe | Compatible endpoints + adapter/switch path | SISCI or endpoint-specific integration | Peer-to-peer support in host, IOMMU, BAR size, drivers |
| Share expensive accelerators between systems | SmartIO Device Lending | NTB-enabled fabric + pooled devices | SmartIO / Device Lending | Supported OS, endpoint reset, SR-IOV, allocation workflow |
| Separate host and I/O over a long distance | Optical transparent or NTB link | Optical adapter pair and compatible cables | Depends on transparent or NTB mode | Distance, fibre type, clock isolation, environmental routing |
Host inventory:
CPU platform, available slot generation, lane width, BIOS and IOMMU behaviour.
Endpoint inventory:
Device type, BAR requirements, driver, peer-to-peer and reset support.
Topology:
Number of hosts, devices, switches, cable paths and expected growth.
Performance:
Payload size, sustained throughput, latency, jitter and CPU-load targets.
Software:
Operating system, kernel version, API preference, safety and cybersecurity constraints.
Deployment:
Distance, power, cooling, rack space, redundancy, service access and lifecycle.
Frequently asked questions
Compatibility checks that determine the final configuration.
Transparent PCIe or NTB—which one should be used?
Use transparent PCIe when one host should own and enumerate remote devices through standard drivers. Use NTB when independent hosts must retain separate address spaces and communicate through managed mappings, memory, DMA, sockets or IP. Multi-host systems normally require NTB or a carefully partitioned multi-root architecture.
Can an existing PCIe device driver be retained?
Usually yes in a conventional transparent expansion design, provided the device and host tolerate the external topology, reset sequence and link behaviour. SmartIO Device Lending can also use standard drivers in supported configurations, but it is not a blanket substitute for transparent expansion.
Can a system mix servers, embedded computers and PXIe?
Yes. Dolphin offers multiple form factors and a cross-platform software model, making heterogeneous systems possible. The exact combination still requires review of PCIe generation, connector, clocking, operating system and supported product pairings.
How far can PCIe be extended?
Distance depends on generation, adapter, connector and cable technology. Copper is generally used for shorter rack or lab connections, while optical solutions support longer separation. The chosen distance must be validated against the exact product and cable combination rather than assumed from the PCIe generation alone.
Does reflective memory require a special memory card?
Dolphin’s PCIe reflective-memory model can use mapped host memory and PCIe multicast rather than relying on a dedicated fixed-size reflective-memory board. This allows the application to define the required memory region, subject to the selected topology and software support.
Can GPUs, FPGAs and NVMe drives be dynamically reassigned?
SmartIO Device Lending is designed for this class of use case. Device compatibility, operating system, reset behaviour, SR-IOV capability and selected hardware must be checked. Dynamic reassignment should be designed as a system workflow, not added after the hardware is fixed.
Is PCIe peer-to-peer automatic?
No. The endpoints, host platform, BIOS, IOMMU configuration, switch topology and drivers must all support the required transaction path. Peer-to-peer feasibility should be tested early with the actual GPU, FPGA, DAQ or storage devices.
What information is needed for a quotation?
Provide host and endpoint models, number of nodes, PCIe generation and lane width, cable distance, transparent or NTB intent, operating system, required APIs, topology, throughput and latency targets, power/cooling constraints and expected production quantity.
Primionics support
From topology review to a deployable PCIe configuration.
Primionics supports Indian engineering teams in selecting the PCIe ownership model, validating host and endpoint compatibility, defining lane and switch topology, choosing the eXpressWare software layer, and preparing a complete hardware, cable and licence configuration.
01
Requirement and topology review
02
Host, endpoint and software compatibility
03
Architecture and BOM proposal
04
PoC or technical validation
05
Deployment and commercial support