Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
28 changes: 28 additions & 0 deletions .claude/knowledge/akida.md
Original file line number Diff line number Diff line change
Expand Up @@ -106,6 +106,28 @@ exception (see above).
(`arrhythmia_classification`). So benchmark all three modes; don't assume.
*Source: measured, AKD1500, 2026-10-09 (#58); numbers in each example's `metrics.json`.*

**[both] Every pass holds at least two layers, and a layer is never split across
passes.**
- A layer writes its outputs into the input buffers of another layer in the same pass;
it can't write to its own input buffers. So each layer shares its pass with the layer
it feeds.
- So HwPr can't rescue a model with one very large layer. That layer and its neighbour
must fit in a single pass together, and the mapper's error names whichever of the two
ran out of NPs. In `akida2/model_zoo/imagenet_mobilenet`, MobileNetV1 1.0's last
pointwise layer takes 22 of the FPGA's 24 NPs, so the classifier that must share its
pass can't map, and the error names the classifier. MobileNetV2 0.75 and 1.0 fail on a
single 21-NP depthwise layer.
- The advantage: even in HwPr, the values passed between layers never leave the chip,
across any number of passes. That is a defining difference from sequences, whose final
values must leave the chip.
- A model that doesn't fit in one sequence can still run on the device if the mapper may
split it into several (`map(device, hw_only=False)`). The three MobileNets above map
onto the FPGA that way, as two or three hardware sequences (akida 2.19, 2026-10-20).
- Akida 1's HwPr (AKD1500) works the same way. The examples above were mapped on Akida 2.

*Source: BrainChip engineering (repo owner), 2026-10-20 (Akida 2) and 2026-10-21
(Akida 1); mapped with akida 2.19 on the Akida 2 FPGA, 2026-10-20 (HUB-19).*

## Platforms and clocks

| | NPs | HwPr | Clock | Notes |
Expand All @@ -118,6 +140,12 @@ exception (see above).
**[both] A node is 4 NPs.** Akida 2 hardware is often described in nodes, so the
6-node FPGA has 24 NPs.

**[Akida 2] On the 6-node FPGA, 3 of the 24 NPs can run dense layers** (one FNP2, with
external weight memory, and two FNP3), and those 3 also serve as convolutional NPs
(CNPs). A large convolution can therefore take them and leave no FNP for a dense layer
in the same pass. *Source: `akida.SixNodesIPv2().mesh`, which matches the FPGA's mesh
(akida 2.19, checked 2026-10-20).*

*Source: akida1/README.md; `brainchip_utils/hardware_utils.py` (`AKIDA_CLOCKS_HZ`);
BrainChip engineering, 2026-10-09 (Akida 2 HwPr, nodes and NP counts).*

Expand Down
2 changes: 1 addition & 1 deletion README.md
Original file line number Diff line number Diff line change
Expand Up @@ -57,7 +57,7 @@ Akida 1 and Akida 2 examples are available today; more Akida 2 models can be fou
| --- | --- | --- | --- |
| **Chip** | AKD1500 | AKD2500 | — |
| **Typical use cases** | Image classification, keyword spotting, object detection | Larger models, higher accuracy targets | Always-on sensing, edge inference |
| **Examples in this repo** | [Image Classification (PlantVillage)](akida1/model_zoo/plant_village) · [ImageNet / AkidaNet](akida1/model_zoo/imagenet_akidanet) · [Visual Wake Words](akida1/model_zoo/vww) · [Keyword Spotting](akida1/model_zoo/speech_commands) · [ECG Arrhythmia](akida1/model_zoo/arrhythmia_classification) | [Visual Wake Words](akida2/model_zoo/vww) · [Image Classification (PlantVillage)](akida2/model_zoo/plant_village) · [Keyword Spotting](akida2/model_zoo/speech_commands) | 🔜 Coming soon |
| **Examples in this repo** | [Image Classification (PlantVillage)](akida1/model_zoo/plant_village) · [ImageNet / AkidaNet](akida1/model_zoo/imagenet_akidanet) · [Visual Wake Words](akida1/model_zoo/vww) · [Keyword Spotting](akida1/model_zoo/speech_commands) · [ECG Arrhythmia](akida1/model_zoo/arrhythmia_classification) | [Visual Wake Words](akida2/model_zoo/vww) · [Image Classification (PlantVillage)](akida2/model_zoo/plant_village) · [Keyword Spotting](akida2/model_zoo/speech_commands) · [ImageNet / MobileNetV1](akida2/model_zoo/imagenet_mobilenet) · [ImageNet / MobileNetV2](akida2/model_zoo/imagenet_mobilenet) | 🔜 Coming soon |

---

Expand Down
43 changes: 43 additions & 0 deletions akida2/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -108,6 +108,49 @@ benchmarks for every mapping, and the steps to reproduce them.
</tr>
</thead>
<tbody>
<tr>
<td rowspan="5"><a href="model_zoo/imagenet_mobilenet">General image classification</a></td>
<td rowspan="5">Classification</td>
<td rowspan="5">ImageNet-1k</td>
<td>8-bit</td>
<td>67.47% top-1</td>
<td>AllNPs</td>
<td align="right">313.992</td>
<td align="right">7.850</td>
<td>MobileNetV1 &alpha;=0.75</td>
</tr>
<tr>
<td>8-bit</td>
<td>62.77% top-1</td>
<td>HwPr</td>
<td align="right">268.127</td>
<td align="right">6.703</td>
<td>MobileNetV1 &alpha;=0.5</td>
</tr>
<tr>
<td>8-bit</td>
<td>47.89% top-1</td>
<td>HwPr</td>
<td align="right">113.428</td>
<td align="right">2.836</td>
<td>MobileNetV1 &alpha;=0.25</td>
</tr>
<tr>
<td>8-bit</td>
<td>63.55% top-1</td>
<td>HwPr</td>
<td align="right">275.338</td>
<td align="right">6.883</td>
<td>MobileNetV2 &alpha;=0.5</td>
</tr>
<tr>
<td>8-bit</td>
<td>58.98% top-1</td>
<td>HwPr</td>
<td align="right">226.843</td>
<td align="right">5.671</td>
<td>MobileNetV2 &alpha;=0.35</td>
</tr>
<tr>
<td rowspan="2"><a href="model_zoo/vww">Person detection</a></td>
<td rowspan="2">Classification</td>
Expand Down
Loading