Domestic FPGAs are often discussed in terms of whether they can enter the main high-speed interface path. This time, Jingwei Qili offers a reproducible engineering path: the PV65 receives dual 4K30 image streams, encapsulates them via HSB IP, and sends them over 10GbE to a Jetson AGX Orin. It does not prove compute capability; it only proves that the bridging stage can run.
The front end of the solution uses dual Sony IMX678 sensors configured for 4K30 RAW12. Each stream enters the PV65 via 4-lane MIPI D-PHY at 1.2 Gbps per lane, with a nominal aggregate of 4.8 Gbps across four lanes. On-chip, the D-PHY receives first, then CSI-2 RX parses the data, which is adapted through SIF into 64-bit AXI-Stream and fed into the HSB FPGA IP. The official PV65 interface capabilities are D-PHY up to 4.5 Gbps per lane, C-PHY up to 3.5 Gsps symbol rate, and LVDS up to 1.5 Gb/s. This configuration remains below the chip's rated limits, indicating that the demo prioritizes link stability rather than saturating the interfaces.
The HSB FPGA IP sits between the sensor stream interface and the host network interface, organizing front-end data into network data and handling host-side device configuration and register access. On the data side, output goes through Host AXI4-Stream to the Ethernet MAC for frame processing, PCS performs physical coding, PMA/SerDes handles high-speed serial transceivers, and after optical-to-electrical conversion, the data reaches the host over a 10GbE cable. The PV65 provides 4 SerDes lanes at up to 12.5 Gbps per lane, sufficient to build a 10GbE link. The control path routes host requests into FPGA control logic, then configures the sensor via board-level I²C. Bridging and inference are explicitly separated; the Jetson handles only the downstream AI workload.
Jingwei Qili is registered in the Beijing Economic-Technological Development Area, with headquarters and R&D center in Yizhuang and subsidiaries in Shanghai and Shenzhen. Officially, it is described as one of the earlier companies to enter independent R&D, volume production, and batch sales of general-purpose FPGAs. In this solution, the PV65 handles interface reception and data movement, not AI inference, nor the sensor itself. For board engineers, the value lies in whether a combination of multi-standard PHY, SerDes, and Ethernet IP can be assembled on a single domestic device; the real questions are HSB IP licensing terms, toolchain maturity, and long-term supply, none of which are provided in the source material.
For domestic teams building high-speed vision boards, this link provides a comparable domestic option: front-end MIPI CSI-2 reception, on-chip AXI-Stream organization, and 10GbE egress, with stage divisions consistent with mainstream solutions. Substitution mainly involves changing PHY configuration and IP interfacing. Adoption depends on whether HSB IP can be obtained, the level of support for timing closure and debugging from domestic toolchains outside Vivado, and PV65 volume pricing and lead times—none of which are disclosed in the source material. In the short term, it is more suitable as an alternative for bridge boards rather than a direct replacement for imported solutions already in volume production.
Daily Briefing · Weekly Digest · Monthly Deep Dive