Is there any plan to have PYNQ support versal cards and boards?

Community corner
1

Versal board is getting more and more popular, and it is 2026 now. Is there any plan to have PYNQ support versal cards and boards?

2

Hi @zzzhhh

Welcome to the PYNQ community!

This is a question that has come up before, and is something we are considering, but it would be good to understand the wider community interest in PYNQ support for Versal.

For example,

  • What are the applications/use-cases?
  • What Versal devices/boards would users be interested in (Edge, Core, RF, etc.)?
  • What Versal features would users like to see supported?

If you (and the wider community) would like to contribute to this conversation, then feel free to reply to this thread, or DM me privately.

3

Hi, @joshgoldsmith,

Thank you for the reply. As a researcher and developer heavily invested in the AMD/Xilinx ecosystem, I believe porting PYNQ to Versal is not just a natural progression; it is a critical necessity for the adoption of ACAP architecture in the wider algorithm and research community. Below are my answers to your inquiry, focusing on use cases that are currently bottlenecked by the complexity of the standard Vivado/Vitis flow.

- My core argument: complexity of and accessibility to Versal

The transition from Zynq to Versal is a massive leap in complexity. We now have to manage Arm cores, PL, AIE, and the NoC altogether.

For algorithm researchers and data scientists, the barrier to entry for Versal is currently high. PYNQ was the “killer app” for Zynq because it allowed software-defined control of hardware. Versal needs this control even more. Without PYNQ, the AI Engines are effectively locked behind a steep learning curve of C++ programming and complex Vitis linking, alienating the Python-native AI community.

- Applications and use-cases
As with many in the modern AI field, my current work focus aligns perfectly with Versal’s hardware, but we need the rapid prototyping capability of PYNQ to make it viable. Specifically, I am working on Embodied AI for robotics. Intelligent robotic systems require acceleration of advanced AI algorithms and complex control to respond to dynamic environment. In my project, one of the neural networks in the AI model requires massive parallelism and custom message-passing schemes that don’t fit standard GPU architectures very well. The Versal AIE array is theoretically perfect for this. Another neural network requires high-speed interaction with the host. The PYNQ will play an important role in developing this robotic system: 1) We need to be able to define the graph structure in Python (e.g., PyTorch Geometric) and offload the compute-heavy node aggregation to AIE overlays via PYNQ, dynamically feeding data through the NoC. 2) A Jupyter-based environment allows for real-time visualization of model training, giving immediate feedback for fine-tuning weights. 3) We can use PYNQ to offload the heuristic computation to the PL and AIE while managing the high-level pathing logic with Python in the PS cores.

- Versal devices/boards
I have access to both VCK5000 and V80 for AI model training in a data-center environment. In robot, we use a Versal Edge board for AI inference to make use of its larger hardware resources and higher computation performance per Watt than Zynq.

- Specific features of Versal requested to be supported in PYNQ
To make PYNQ on Versal truly useful, we need more than just PL control. I wish PYNQ could support the following Versal features:

  1. AIE control: Python bindings to load graph binaries (.xclbin) and to control runtime parameters of the AI Engines directly from a notebook.
  2. NoC memory management: An improvement of pynq.allocate that is NoC-aware, allowing us to easily pin buffers to specific memory banks to optimize bandwidth for the AIEs.
  3. Composable overlays for AIE: Pre-built overlays where the PL acts as a generic data mover feeding custom AIE graphs. This would allow software users to swap AIE logic without touching Verilog.

To sum up, Versal is the hardware of the future for heterogeneous computing, but Python is the language of the present AI community. PYNQ is the only bridge that connects them effectively. Bringing PYNQ to Versal would empower researchers to stop fighting the build tools and start innovating on the algorithm and architecture (for smarter AI agents for example).

Thank you for considering this roadmap.

1 Like