System Requirements#
Hardware Requirements#
The following table lists AMD Vitis™ AI developer workstation hardware requirements:
Component |
Requirement |
|---|---|
Host Machine |
The recommended host environment is a 64-bit Linux machine equipped with at least 20 cores, 64 GB of RAM, and 100 GB of free disk space, running either Ubuntu 22.04.4 LTS or 24.04.2 LTS. While Ubuntu 24.04.2 LTS is compatible with Vitis and Vivado tools, it is not currently supported by Yocto. |
SD Card |
SD card with a minimum capacity of 32 GB. |
Target Board |
VEK385 RevA or VEK385 RevB card with pre-production silicon. |
Connectivity |
VEK385 target board connected to the host machine with an Ethernet cable (optional) and a UART USB cable. The cable (Type A to Type C) is provided with the VEK385 board for connectivity. Type A is connected to the host and Type C is connected to the board. |
Note
You can connect the VEK385 target board to a host machine for convenient file transfers, for example, the model cache directory and overlay files. However, you do not require a network connection to run inference.
Software Requirements#
Component |
Requirement |
|---|---|
Host OS |
Ubuntu 24.04.2 LTS, Ubuntu 22.04.4 LTS. While Ubuntu 24.04.2 LTS is fully supported by AMD Vitis and Vivado tools, it is not currently supported by Yocto. For Yocto-based workflows, Ubuntu 22.04.4 LTS is recommended. |
Vivado |
Version 2026.1 |
Vitis |
Version 2026.1 |
Python |
Version 3.12 |
PyTorch |
Version 2.5.1 |
ONNX |
Version 1.16.1 |
ONNX Runtime |
Version 1.28.0 |
Docker |
Version 28.5.1 |
Quark (included in the Docker image) |
Version 0.12 |
Software Packages |
Includes binaries for PDI, Linux images, and models |
ONNX Opset Support#
The Vitis AI Compiler supports a defined range of ONNX opset versions. Use the following guidance when you export or convert a model for compilation.
Opset Version |
Support Level |
Notes |
|---|---|---|
11 - 20 |
Fully supported |
Opset 20 is the recommended version. |
21 - 22 |
Partially supported |
Some operators may not be compiled. Validate your model before deployment. |
Above 22 |
Not supported |
Re-export or down-convert the model to a supported opset. |
Support Details#
Opsets 11 through 20 are fully supported. Opset 20 is recommended for best compatibility and performance.
Opsets 21 and 22 are partially supported. Models that use these opsets may compile, but some operators are not guaranteed to be supported.
Opsets above 22 are not supported. Models exported with these opsets must be re-exported or converted to a supported opset before compilation.
Note
If your model uses an unsupported opset, re-export it from the source framework by targeting opset 20, or use the ONNX version converter to down-convert the model.
Note
Docker versions up to 28.5.1 are verified as compatible. Newer versions might work, but AMD has not tested them with Vitis AI.
License Requirements#
The following licenses are required before proceeding:
AIE-MLv2 Compiler: Required for compiling ONNX models with the Vitis AI compiler in the Docker container. Refer to Obtaining License for instructions on obtaining this license.
Vivado ML Enterprise Edition: Required for building the Vitis AI reference design.