PS-PL Connections - UG1192

AMD Design Conversion for Altera FPGAs and SoCs Methodology Guide (UG1192)

Document ID
UG1192
Release Date
2025-07-15
Revision
3.0.1 English

The ports connecting the PS and PL (referred to as HPS and FPGA in Altera terminology) are similar to those in the Zynq UltraScale+ MPSoC and Versal devices with Cortex-A72 based PS, but with some differences. AMD devices include general-purpose AXI ports, high-performance AXI ports, and an accelerator coherency port (ACP), among others differences. Altera devices use the FPGA-to-SDRAM interface, FPGA-to-HPS bridge, HPS-to-FPGA bridge, and the lightweight HPS-to-FPGA bridge. See the following tables.

Table 1. Zynq UltraScale+ MPSoC vs. Agilex 5: PS-PL Connections
HPS ↔ Fabric Agilex 5 Equivalent Zynq UltraScale+ MPSoC Designing to Manage the Differences
FPGA-to-SDRAM bridge (F2SDRAM) 1x 256-bit AXI4 fabric access to DDR memory

Lower latency than F2SOC support

Non-coherent with the HPS

4x 128-bit PL to PS-FPD Improved throughput: The four 128-bit paths can be used in parallel to achieve higher throughput compared to a single 256-bit path. This can be managed by distributing the data across the four paths, ensuring balanced load and avoiding bottlenecks.
FPGA-to-HPS bridge (F2H or F2SOC) 1x 256-bit ACE-Lite5 fabric access to DDR, OCRAM, and HPS peripherals

Coherent and non-coherent transactions

Cache stashing into L1, L2, and L3 caches

4x 128-bit PL to PS-FPD (HPC ports)

128-bit AXI PL to PS-LPD for OCM and PS peripherals

128-bit ACP exclusively for L2 cache

Coherency management: Use the multiple 128-bit paths to separate coherent and non-coherent transactions, reducing contention and improving performance. Cache stashing can be managed by directing specific transactions to the ACP for L2 cache.
HPS-FPGA bridge (H2F) 1x 128-bit AXI4 port from HPS to fabric for high bandwidth access 128-bit AXI4 PS-LPD to PL and 2x 128-bit PS-FPD to PL for non-coherent and coherent transactions Bandwidth optimization: Use the multiple 128-bit paths to segregate high-bandwidth and low-latency transactions, ensuring efficient data transfer and reducing the risk of congestion.
Lightweight HPS-to-FPGA bridge (LWH2F) 1x 32-bit AXI4 port from HPS to fabric for low bandwidth access to soft IP control and status registers 128-bit AXI4 PS-LPD to PL  
Peripheral fabric ports Using fabric, I/Os for HPS peripherals EMIO  
  Not available 128-bit ACE from PL to PS for fully coherent transactions

Additional AXI4 resource: PL master with cache coherent transactions

Table 2. Versal Device Cortex-A72 Based PS vs. Agilex 5: PS-PL Connections
HPS ↔ Fabric Agilex 5 Equivalent Versal Device Cortex-A72 Based PS Designing to Manage the Differences
FPGA-to-SDRAM bridge (F2SDRAM) 1x 256-bit AXI4 fabric access to DDR memory

Lower latency than F2SOC support

Non-coherent with the HPS

NoC for DDR memory

The NoC-based memory access in Versal devices provide better scalability and efficiency compared to direct AXI4 connections.

It helps optimize bandwidth across multiple masters without congestion issues (programmable connectivity/NoC QOS settings)

FPGA-to-HPS bridge (F2H or F2SOC) 1x 256-bit ACE-Lite5 fabric access to DDR memory, OCRAM, and HPS peripherals

Coherent and non-coherent transactions

Cache stashing into L1, L2, and L3 caches

NoC for DDR memory

128-bit AXI4 PL to PS-LPD for OCM and PS peripherals

128-bit ACP exclusively for L2 cache

NoC-based access provides better routing and scalability compared to direct fabric-based access.

Ability to separate coherent and non-coherent transactions

HPS-FPGA bridge (H2F) 1x 128-bit AXI4 port from HPS to fabric for high-bandwidth access 128-bit AXI4 PS-LPD to PL and PS-FPD to PL for non-coherent and coherent transactions

Versal devices provide both coherent and non-coherent access paths.

Coherent transactions to reduce redundant memory operations.

Lightweight HPS-to-FPGA bridge (LWH2F) 1x 32-bit AXI4 port from HPS to fabric for low bandwidth access to soft IP control and status registers 128-bit AXI4 PS-LPD to PL Versal devices provide much higher bandwidth compared to Agilex
Peripheral fabric ports Using fabric, I/Os for HPS peripherals EMIO  
  Not available 512-bit ACE from PL to PS for fully coherent transactions

Additional AXI4 resource