PCIe400 generic readout board qualification test G. Vouters Abstract The PCIe400 is a generic board for high-throughput data acquisition system s in high energy physics experiments. Its purpose is to interface up to 48 bidirectional links, supporting custom protocols at 1 to 26 Gbit / s \mathrm{Gbit}\text{/}\mathrm{s} , to modern commercial back-end links providing 400 Gbit / s \mathrm{Gbit}\text{/}\mathrm{s} bandwidth. It also targets clock distribution with phase determinism below 10 ps peak-to-peak. It has been designed for LHCb LS3 enhancement upgrade with experimental features to prepare LHCb Upgrade II, foreseeing an aggregated throughput of 200 Tbit / s \mathrm{Tbit}\text{/}\mathrm{s} . However, its versatility allows it to be used in several experimental environments. The board embeds Altera’s flagship Agilex 7 M-series FPGA with a PCIe Gen 5 interface and an experimental QSFP112 serial interface. We present the results of qualification tests performed on prototype boards and the challenges encountered to meet specifications. Section 1 describes board-level validation, including power-up behavior and peripheral access. Section 2 focuses on high-bandwidth interface qualification through BER measurements. Finally, Section 3 investigates phase determinism in Agilex transceivers, a key requirement for precise clock distribution. 1 Early stage validation The board is designed around Altera’s Agilex 7 M-series [ 1 ] with 1.3 million ALM 1 1 1 Adaptive Logic Module running at up to 1 \mathrm{GHz} and 32 GB of HBM2e 2 2 2 High Bandwidth Memory . It is a double width, 268 \mathrm{mm} long PCIe add-in card. It features up to 48 bidirectional links for custom protocols at 1 to 26 \mathrm{Gbit}\text{/}\mathrm{s} each, a PCIe Gen 5 x16 connector and a 400GbE interface for back-end systems. Finally, it can manage various Time Fast Control protocols , such as TTC-PON [ 2 ] and White Rabbit [ 3 ] . A block diagram and a photo of the board are given in Figure 1 . Figure 1 : (left) Block diagram of the PCIe400 board highlighting the FPGA as the central element interfacing PCIe, 400GbE and custom high-speed links. (right) Top view of the PCIe400 prototype illustrating the high component integration density. 1.1 Powering the board The board features 23 power rails, with strict requirements on voltage accuracy and ripple noise within \pm 0.5\% . Using a custom-made coaxial probe coupled with a DC-blocking capacitor [ 4 ] , all voltage rails pass requirements with an idle FPGA consuming 24 \mathrm{W} . During PCIe enumeration, the board must power-up within 100 \mathrm{ms} [ 5 ] , and the FPGA within 10 \mathrm{ms} to meet Altera’s CvP (Configuration via Protocol) link-up requirements [ 6 ] . The power-up time initially exceeded specifications, while the FPGA started in 7.5 \mathrm{ms} . This was caused by an unexpected 100 \mathrm{ms} fixed delay from the hotswap controller and a schematic error preventing a DC/DC converter to start reliably. After a rework correcting the latter, the power-up time was reduced to 134 \mathrm{ms} . A revision to replace the hot-swap controller with a no-delay version would achieve compliance with 60% margin on specifications. Finally, the PCIe test server includes a 500 \mathrm{ms} internal reset delay, relaxing the requirements and enabling further tests. 1.2 Peripheral access The board can be controlled through PCIe and USB/JTAG for debugging. These connections allow configuring peripherals and monitoring board status. External components and internal gateware module registers are accessible through a common Avalon Memory-Mapped bus. A test suite, built upon our framework, pytest [ 7 ] and polars [ 8 ] , validated all peripheral access through more than 400 compliance tests executed in less than a minute. The framework is generic and organized in four layers as described in Figure 2 . There are three major test suites: a Functional Test comparing sensor values to a reference , a Collect Test to monitor a selection of parameters for a given time and a Console for interactive debugging. Figure 2 : Software architecture of the PCIe400, showing the layered organization used for functional validation, monitoring, and interactive debugging. 2 High bandwidth interface One figure of merit for assessing high-bandwidth interface integrity is the bit error rate (BER). Measurements were performed on both backend interfaces using pseudo random bit patterns (PRBS). 2.1 PCIe interface The PCIe interface is based on previous developments for the generic readout board PCIe40 [ 9 ] . Thus , the tests were performed in 2x8 bifurcated mode , as data streams could always be split in two in previous applications and because it simplifies gateware implementation. The tests , illustrated in Figure 3 are performed on a PCIe Gen 4 server only, while waiting to get access to a Gen 5 server. Figure 3 : PCIe tests performed on a PCIe Gen 4 server in 2×8 bifurcated mode, demonstrating correct enumeration and DMA transfers with a BER below 10^{-15} at 95% confidence level on both interfaces. 2.2 4x100 Gbit/s interface The Agilex F-Tile FHT transceiver supports 116 \mathrm{Gbit}\text{/}\mathrm{s} PAM4 per lane and includes a 400GbE hard IP enabling the integration of a high-bandwidth network interface on-board. The main challenge lies in the hardware link between the FPGA and the QSFP112 connector, separated by approximately 50 \mathrm{mm} [ 10 ] . The test setup uses an FS QSFP112-SR4-400G module in external loopback and Altera’s reference design featuring RS(544, 514) FEC with 15 correctable symbols. As shown in Figure 4 , seven hours of PRBS testing yields excellent signal integrity, with a maximum of two corrected symbols per codeword. Figure 4 : PRBS-31 bit-error-rate measurement on the 4x100 Gbit/s interface. Seven hours of continuous testing show no uncorrected errors, corresponding to a BER below 1\cdot 10^{-15} at 95% confidence level. 3 Phase determinism According to the datasheet, Agilex FGT transceivers exhibit a non-deterministic lane-to-lane output skew of up to \pm~2 unit interval (UI) + 300 \mathrm{ps} [ 11 ] . This translates to jumps of 500 \mathrm{ps} at 10 \mathrm{Gbit}\text{/}\mathrm{s} and prevents a straightforward solution to distribute phase-deterministic clocks to the front-end. Tests were intentionally performed on an Agilex DK-SI-AGI027FA development kit in order to isolate the intrinsic phase behavior of the transceiver from board-level effects. [ 12 ] 3.1 Principle and set-up The design implements four channels running at 10.24 \mathrm{Gbit}\text{/}\mathrm{s} and a 32b fixed pattern. The phase is measured internally using a 1 \mathrm{ps} resolution DDMTD [ 13 ] as described in Figure 5 . Figure 5 : Test setup implemented on the development kit to study transceiver phase behavior. Four serial channels operate at 10.24 Gbit/s, with phase differences measured internally using DDMTD. The tx_clkout signal, corresponding to clock generated by the transceiver logic, can not be used to measure the phase of the output serial stream due to non-deterministic internal alignment. Instead, the transceiver operates in serial-loopback and the recovered clock (rx_clkout) measures the image of the transmitted signal . Figure 6 illustrates the receiver determinism test: after intentionally shifting the transmitted data pattern by one bit and repeating the receiver synchronization, the phase of the recovered clock is displaced by exactly one unit interval, as expected. Figure 6 : Measured phase shift of the recovered clock rx\_clkout following a one-bit shift of the transmitted data stream and repeated receiver synchronization. The observed unit-interval phase steps confirm deterministic receiver behavior after synchronization. 3.2 Tx reset stability The Figure 7 (left) shows that successive transmitter resets produce quantized phase shift of \pm 3\,\mathrm{UI}\allowbreak~(\approx 300~\mathrm{ps}) . With increased statistics, a bimodal distribution appears around each bit position, separated by \approx 20~$\mathrm{ps}$ . A test with 500 Rx synchronizations and no Tx reset reveals that this double-peak pattern persists, as in Figure 7 (right). It indicates that the phase uncertainty results from the receiver path and likely from the Clock Data Recovery circuit locking on two stable positions. Thus, it does not impact the transmitter serial stream but only the phase monitoring system. This can be mitigated by the selecting the most probable position. Figure 7 : (left) Quantized phase shifts of the recovered clock observed after successive transmitter resets, showing jumps limited to a few unit intervals. (right) Phase distribution obtained from 500 receiver synchronizations without transmitter reset, revealing two stable locking positions of the receiver clock recovery. Finally, we assess the inter-reset stability of the transmitter serial stream on four channels. The bimodal distribution is mitigated by defining a \pm 10~$\mathrm{ps}$ filter window around the most probable phase position. The result, in Figure 8 , shows a stability of 6 to 8 \mathrm{ps} peak-to-peak over four channels after 500 resets during 12 hours demonstrating sub-10 ps stability across channels. Figure 8 : Phase stability of the recovered clock measured over four serial channels after 500 transceiver resets. The distributions show a peak-to-peak phase variation of 6–8 ps. 4 Summary In the process of qualifying the generic readout board PCIe400, we developed a versatile software framework that enables efficient testing of complex boards from debugging to production. Extensive simulation work during design allowed the implementation of errorless 4x100 \mathrm{Gbit}\text{/}\mathrm{s} PAM4 links and PCIe interface. A phase - determinism proof of concept on four channels show s that Agilex FGT transceivers are compatible with a 10 \mathrm{ps} peak-to-peak phase determinism precision. Additional tests are underway to qualify all PCIe400 features. In particular, further studies on clock distribution with phase determinism are required to evaluate stability across gateware compilations, temperature variations and at a system level.