Metis PCIe: Voyager SDK Tiling Error on 1504×1504 Input (L1 Constraint Exceeded)

Hi ​@dany! Good question. Looks like you’re reading the compiler output right.

That l1_constraint line refers to the per-core on-chip memory on the Metis AIPU. The compiler automatically tries to fit within that limit, but it looks like the tile still overflows the L1 buffer.

I wonder if a slightly lower input resolution (e.g. 1280×1280) could help with this?

Hi ​@Spanner,

Thanks for your reply. Yes, a lower input resolution works fine. I was successfully able to deploy a model with a custom input resolution of 1504×1088.

We anticipate to use a higher resolution (e.g., 1504×1504) in the near future. Therefore, I am checking the available possibilities for achieving this on the Metis PCIe card.

Hi ​@dany,

Thank you for reaching out and bringing this issue to our attention.

The error you're encountering is unusual and warrants further investigation. The compiler should successfully compile your model within the L1 memory limit, but something about the model or operator is causing the compiler to get confused. It could also be that the model shapes limit the compiler's ability to apply the transformations required to fit the operator to the hardware constraints.

To help diagnose the problem, would it be possible for you to provide details on the operator (type, input/output shapes) and/or information on the model (ONNX file or architecture details)?

This will enable us to identify the root cause more efficiently.

Best,
Fabian

Hi Fabian,

Thank you for the follow up.

I used the Yolo11n model architecture and the yolo11n-coco-onnx.yaml file for deployment. The input ONNX model is yolo11n_1504x1504.onnx, and the input tensor shape is [1, 3, 1504, 1504]. 

The only functional modifications made to this file were setting the ONNX weight path and fixing the input tensor shape, as shown in the snippet below:

models:
  yolo11n-coco-onnx:
    class: AxONNXModel
    class_path: $AXELERA_FRAMEWORK/ax_models/base_onnx.py
    weight_path: $AXELERA_FRAMEWORK/customers/test_coco/yolo11n_1504x1504.onnx
    # weight_url: https://media.axelera.ai/artifacts/model_cards/weights/yolo/object_detection/yolo11n.onnx
    # weight_md5: d00108b1a46170499d67a80825531d26
    task_category: ObjectDetection
    input_tensor_layout: NCHW
    input_tensor_shape: [1, 3, 1504, 1504]
    input_color_format: RGB
    num_classes: 80
    dataset: CocoDataset-COCO2017
    extra_kwargs:
      cal_seed: 129
      num_cal_images: 100
      compilation_config:
        quantization_scheme: per_tensor_min_max
        ignore_weight_buffers: false