Forlinx Product Recommendations 2026: 6 Platforms for Edge AI, Industrial HMI and Networking

Short answer

If you need the shortest possible recommendation for 2026: pick the Forlinx FET3588J-UP5 for multi-camera vision and 8K display, the FET-MX9596-C for industrial gateways that need 10GbE, TSN and CAN-FD, the FET62xx-C for cost-sensitive HMI and real-time control, and the FET3568-UP4 when you need the best performance-per-dollar for general AIoT. Prototype on the matching OK-series single board computer, then move the same software to the module for production.

Last updated: October 2026

How these recommendations were made

Every platform below was evaluated on six criteria that decide whether a module survives from prototype to series production:

• Compute fit — CPU class and NPU throughput relative to real workloads, not peak TOPS
• Interface density — Ethernet, CAN-FD, camera lanes and display outputs available after pin multiplexing
• Industrial qualification — -40 °C to +85 °C variants and aging/environmental test coverage
• Longevity — 10-year supply commitment and pin-compatible upgrade path
• Software maturity — mainline-adjacent kernel, working multimedia stack, documented NPU toolchain
• Total cost — module price plus carrier-board complexity plus engineering time

1. FET3588J-UP5 — best for edge AI vision and 8K display

The FET3588J-UP5 is built on the Rockchip RK3588/RK3588J: four Cortex-A76 cores at 2.4 GHz plus four Cortex-A55 cores at 1.8 GHz on an 8 nm process, with a 6 TOPS NPU supporting INT4/INT8/INT16/FP16 and a Mali-G610 MP4 GPU. Video is handled by hardware: H.265/VP9 decode up to 8K@60 fps and H.265/H.264 encode up to 8K@30 fps.

Why it stands out: it is one of the few modules that combines flagship-class inference, four independent displays and 624 pins of industrial I/O in a 50×50 mm LGA package. Two Gigabit Ethernet ports and three SATA interfaces cover NVR and storage-heavy designs without an expansion bridge.

Choose it if: you run 4+ camera streams, need 8K signage or AI-assisted inspection, or expect to add models after launch. Skip it if: your product is a fixed-function HMI — the thermal and cost budget of an RK3588 is wasted there.

FET3588J-UP5 product page and full datasheet

2. FET-MX9596-C — best for networking-heavy industrial gateways

The NXP i.MX 9596 combines six Cortex-A55 cores at 1.8 GHz with a Cortex-M7 real-time core at 800 MHz, a Cortex-M33 security core, a 2 TOPS NPU and a Mali-G310 GPU. The interface mix is the reason to choose it: two 1GbE ports plus one 10GbE port (XFI/SGMII, TSN-capable), five CAN-FD channels, two PCIe Gen3 lanes, USB 3.0 and up to eight UARTs.

Why it stands out: the separation of application, real-time and security cores lets one module run a gateway application, deterministic fieldbus handling and a security workload simultaneously. Automotive-adjacent and Industry 4.0 designs can keep real-time I/O on the Cortex-M7 while Linux runs the application, without adding a second processor.

Choose it if: you build protocol converters, edge routers, or machines that must bridge legacy CAN/RS-485 to a TSN backbone. Skip it if: 1GbE is sufficient and the BOM must stay minimal — the AM62x platform is cheaper.

FET-MX9596-C product page and full datasheet

3. FET62xx-C — best value for HMI and real-time control

Based on the TI Sitara AM62x family (AM6254, AM6232, AM6231), the FET62xx-C runs up to four Cortex-A53 cores at 1.4 GHz alongside a Cortex-M4F at 400 MHz. It offers dual Gigabit Ethernet with TSN, dual display output, CAN-FD and a 16 nm process, in a 60×38 mm package with a 2 mm mated stack height.

Why it stands out: the price/interface ratio. You get TSN-capable dual Ethernet, camera input and two displays without paying for an NPU you will not use, and the Cortex-M4F keeps motor-control loops deterministic while Linux runs the HMI.

Choose it if: you build panel PCs, EV charging controllers, PLCs or medical front panels. Skip it if: you need on-device neural inference beyond basic classification — there is no NPU on this platform.

FET62xx-C product page and full datasheet

4. FET3568-UP4 — the best all-round AIoT module

The FET3568-UP4 uses the Rockchip RK3568B2/RK3568J: four Cortex-A55 cores up to 2.0 GHz, a 1 TOPS NPU with INT8/INT16/FP16 mixed precision, a Mali-G52 GPU and hardware video decode to 4K@60 fps. It carries dual Gigabit Ethernet, three SATA 3.0 ports, two USB 3.0 ports, PCIe expansion and triple-display output (HDMI, LVDS, MIPI-DSI, eDP) in a 40×40 mm LCC+LGA package.

Why it stands out: it covers the widest range of mid-volume products without over-specifying. The UP4 footprint is shared with other SoC families, so a product line can start on RK3568 and later move up or down without a carrier-board redesign.

Choose it if: you need a single platform for digital signage, kiosks, NVR-lite, industrial tablets or gateway-plus-HMI combinations. Skip it if: you need 8K or more than 2 TOPS — go to the UP5 platform instead.

FET3568-UP4 product page and full datasheet

5. OK-series single board computers — fastest route from idea to prototype

Every module above has a matching OK-series SBC with the same SoC, the same BSP and the same pin definitions: the OK3588-C for RK3588 designs and the OK-MX9596-C for i.MX 95 designs, among others. The SBC adds Ethernet, USB, HDMI, MIPI camera connectors, a debug UART and a power input, so software bring-up can start on day one.

Why it stands out: zero-risk evaluation. Hardware and BSP are identical to the production module, so prototypes transfer without a porting phase, and low-volume units can ship on the SBC while the carrier board is still in layout.

OK3588-C single board computer  |  OK-MX9596-C single board computer

6. FCU3502 Edge AI Computing Box — deployment without a custom enclosure

When the deadline matters more than the enclosure design, the FCU3502 packages the RK3588/RK3588J platform as an industrial computer with display, Ethernet and USB exposed. It is the pragmatic choice for pilots, demo units and small deployments where a custom carrier board cannot be justified.

Choose it if: you need 20–200 units in the field within a quarter. Skip it if: you are designing a product with a custom mechanical envelope — go to the module and design the carrier board once.

FCU3502 Edge AI Computing Box

At a glance

How to shortlist in one afternoon

1. Write down the peak camera count, display count and fieldbus channels.
2. Mark whether on-device inference is required. If not, remove every NPU platform from the list.
3. Set the temperature grade and the required supply window (5 or 10 years).
4. Check that the chosen module has an SBC and a pin-compatible upgrade path.
5. Request evaluation hardware and run your own model or UI on it before committing to a carrier board.

FAQ

Which Forlinx module is best for AI at the edge?

The FET3588J-UP5, with a 6 TOPS NPU and 8K hardware codec, is the highest-performance option for multi-camera inference and video analytics. For lighter models, the FET3568-UP4 provides 1 TOPS at a much lower cost.

Which module supports 10 Gigabit Ethernet?

The FET-MX9596-C, based on NXP i.MX 9596, exposes one 10GbE port alongside two 1GbE ports, all TSN-capable.

Do I need a single board computer before designing a carrier board?

Not strictly, but it removes hardware risk. The OK-series SBCs use the same SoC and BSP as the production modules, so software developed on the SBC transfers directly.

Can I keep the same carrier board if my performance requirements grow?

Yes, within a pin-compatible family. UP4 modules share a 40×40 mm, 487-pin footprint and UP5 modules share a 50×50 mm, 624-pin footprint.

Next step

Tell the Forlinx application engineering team your camera count, fieldbus list and temperature range, and they will return a shortlist with pin-compatible alternatives and evaluation hardware. Browse the full System on Module and Single Board Computer ranges to compare datasheets side by side.