FPGA AI Suite: Version 2024.3 Release Notes

ID 772497
Date 12/05/2024
Public

A newer version of this document is available. Customers should click here to go to the newest version.

4. Known Issues and Workarounds

FPGA AI Suite Version 2024.3 has the following known issues:
  • When the input layout transform is enabled, the area estimates for some IP architectures can be too low.
  • Building the Arria® 10 SoC design example on a Red Hat Enterprise Linux system requires Red Hat Enterprise Linux 8.10.
  • The area model reports the wrong DSP configuration with Agilex™ 5 architectures.

    The area model reports the DSP configuration being "Sum of 4 9x9" when it is a DSP tensor mode. This reporting results in a "PTC Import Warning" when you use a .ptc file generated from an Agilex 5 architecture.

  • The AGX7_Streaming_Ddrfree_Multilane architecture file does not have a pool aux module. This lack of the pool aux module can produce dla_compiler errors on graphs that require pooling, for example, ResNet18 and ResNet50.
  • If you use the hardware layout transform, then the max_feature_height parameter must be greater than or equal to the channel dimension of your graph. This limitation is due to a known issue with the parameter checks that causes the the input graph channel dimension to be compared against the max_feature_height instead of the max_channels value.

Streaming Flow Known Issues and Workarounds

  • Only single-output graphs are supported.
  • MobileNet-v2 is known to fail in the streaming flow when no external memory is used.

Multilane

  • The num_lanes parameter takes only the values 1, 2, or 4.
  • When an architecture has a num_lanes parameter value greater than 1, the architecture has the following restrictions:
    • All k_vector and c_vector parameter values in the architecture must be the same.
    • The softmax auxiliary module is not supported.
  • Multilane exhibits low performance when used with DDR when a graph has residual connections.