Unsupported operators on Coral Edge TPU, RKNN, Hailo and Kneron
Edge NPU compilers reject some ops outright and quietly push others to the CPU, often after a long compile. These are the rules ModelVisio's compiler pre-flight checks in your browser, taken from each vendor's documented limits.
Check your model now ONNX, TFLite, PyTorch, Keras, Caffe, OpenVINO and more. Problems are marked on the graph.
Rules by target
| Rule | Coral Edge TPU | Rockchip RKNN (RK3588) | Hailo-8 (DFC) | Kneron KL720 |
|---|---|---|---|---|
| Dynamic input shape (any non-batch dim unknown) | ✕ error | ! warning | ✕ error | ✕ error |
| Data-dependent ops (NMS, NonZero, TopK, Loop/If, Scatter…) | ✕ error | ! warning | ✕ error | ✕ error |
| Ops outside the Edge TPU op list (34 ops) | ✕ error | not flagged | not flagged | not flagged |
| MatMul of two activations (attention-style) | ✕ error | not flagged | not flagged | not flagged |
| Softmax not on a 1-D tensor of ≤ 16,000 elements | ✕ error | not flagged | not flagged | not flagged |
| Output with more than 3 non-unit dimensions | ✕ error | not flagged | not flagged | not flagged |
| Resize / Upsample | ! warning | not flagged | not flagged | not flagged |
An error means the compiler rejects the op or the device can't run it. A warning means it may run on the CPU or needs an extra build option. "Not flagged" means ModelVisio has no rule for it, not a guarantee that it compiles.
Static input shapes
Every one of these compilers builds a fixed schedule, so every tensor size has to be known at compile time. A model exported with a dynamic height, width or sequence length fails on Coral, Hailo and Kneron. RKNN builds for fixed shapes or a preset list (dynamic_input). Re-export with a static input size, for example 1×3×640×640. The batch dimension is fixed at export and is not flagged.
Data-dependent ops
These ops decide their output size or control flow from the input values, so no static NPU schedule exists for them. Cut the graph before them and do the step in post-processing on the host. Hailo handles NMS in its own post-process.
nonmaxsuppressionnonmaxsuppressionv4nonmaxsuppressionv5nmsmulticlassnmsmulticlassnms3nonzerouniquetopktopkv2loopifscanwhilescatterndscatterelementsscatterdetectionoutputCoral Edge TPU
The Edge TPU compiler maps the graph only up to the first unsupported op. Everything after it runs on the CPU, even ops the TPU supports, so one early unsupported op can cost most of the speed-up. After ONNX or Keras is lowered to TFLite, these ops become DIV, POW, EXP, LOG, SQRT, GATHER, CAST and similar ops outside the supported list:
divrealdivfloordivpowexplogsqrterfgelugathergatherv2gatherndgatherelementscasteinsumcumsumlayernormalizationlayernorminstancenormalizationinstancenormgroupnormalizationgroupnormmishsoftpluseluseluwhereselectselectv2argminmodsincosreciprocalShape and pattern limits
- Softmax runs on the TPU only on a 1-D tensor of at most 16,000 elements.
- MatMul between two activations (attention-style batch matmul) is not supported. Only fully connected layers with constant weights map.
- Tensors may have a size above 1 only in their 3 innermost dimensions.
- Resize and Upsample may stay on the CPU depending on input and output size, to avoid precision loss. Check the compiler log.
Auto-fixes in ModelVisio
For the common cases the app can rewrite the graph in place, show the new op-support result, and undo it:
- SiLU → HardSwish. Replace SiLU/Swish activations with HardSwish (piecewise-linear). App's estimate: ≈ same accuracy (−0.2% typical) · fuses on NPUs · no Sigmoid+Mul CPU fallback.
- SPPF → parallel SPP. Restructure the cascaded MaxPool chain (SPPF) into parallel pooling branches. App's estimate: Compiles on Coral / Kneron · ≈ +5% latency · identical 13×13 receptive field.
- Resize(nearest) → ConvTranspose. Replace nearest-neighbour Resize/Upsample with a stride-2 transposed convolution. App's estimate: Runs fully on-chip (no Resize CPU fallback) · adds a small learned upsampler.
- Channel prune 0.75×. Structured channel pruning (keep 75%) on the heaviest layers (> 200K params). App's estimate: −25% params/FLOPs on pruned layers · ≈ −0.5% mAP · needs a short fine-tune.
The fixes change the in-browser graph so you can see the effect on compatibility and the estimated latency. Apply the same change in your training or export code before you compile.
How the checks work
Op names are normalized across ONNX, TFLite, Keras, Caffe, OpenVINO and other spellings, so stablehlo.dot_general and MatMul are matched the same way. The rules are deliberately conservative: an op ModelVisio doesn't recognize produces no finding. The check can miss a problem but won't invent one. Integer shape arithmetic that converters constant-fold away is skipped. The vendor compiler still has the final word.
Sources: Coral: TensorFlow models on the Edge TPU (supported operations and model requirements); the Rockchip RKNN, Hailo Dataflow Compiler and Kneron toolchain documentation. See also the edge AI accelerator comparison.