Model Explorer
The vision models behind the RVPU pipelines.
A catalog of the depth, detection, and segmentation models the SDK compiles and configures into RVPU pipelines. Until our first Kria demo lands, throughput rows are pending — we publish measured numbers only once we have run them on our own hardware. The one preliminary target lives on the Benchmarks page, clearly labelled.
| Family | Precision | Memory | Status | |||
|---|---|---|---|---|---|---|
SAM ViT-Base Encoder integrated; segmentation/scene pipeline throughput pending our demo. | SAM | 91M | INT8 | 46 MB | —FPS | Pending |
DETR ResNet-50 Compile/config path validated; throughput benchmark pending. | DETR | 41M | FP16 | 82 MB | —FPS | Pending |
DeepLabV3 ResNet-50 Scene parsing / SLAM-assist segmentation. Throughput pending. | DeepLab | 39M | INT8 | 78 MB | —FPS | Pending |
YOLOv8 Medium | YOLO | 25.9M | INT8 | 52 MB | —FPS | Pending |
Depth Anything V2 (Small) Anchor of the dense-3D pipeline. Preliminary target (30+ Hz on Kria KV260) is on the Benchmarks page — pending our own demo. | Depth Anything | 24.8M | INT8 | 25 MB | —FPS | Pending |
YOLOv8 Small | YOLO | 11.2M | INT8 | 22 MB | —FPS | Pending |
YOLOv8 Nano 640×640 input. Backs person/object and thermal–RGB detection pipelines. Throughput pending measurement on our hardware. | YOLO | 3.2M | INT8 | 6 MB | —FPS | Pending |
Stereo matching (real-time) Pairs with monocular depth for dense stereo 3D. Throughput pending our Kria demo. | Stereo | — | INT8 | — | —FPS | Pending |
Methodology
How throughput numbers are derived.
- 01Status legend — measured: a real run on our own Kria hardware. projected: a modelled estimate, not yet measured. pending: the SDK compile/config path is validated but throughput has not been measured on our hardware.
- 02Every throughput row here is pending. We do not publish perception numbers we have not measured ourselves. The single preliminary target — Depth Anything V2 + stereo at 30+ Hz on a Kria KV260 — is on the Benchmarks page and labelled "Preliminary — target."
- 03Pending rows have a working compile/config path through the SDK, so a team can confirm pipeline support today even though the timing run has not landed.
- 04Design partners get earlier visibility into preliminary demo data under NDA as it becomes available.
Don't see your model?
The SDK compiles PyTorch, ONNX, and HuggingFace models into pipeline stages — if you can export a detector, we can probably fold it into a pipeline. Tell us the pipeline you need and the target platform and we will get it on the explorer.