diff --git a/Benchmarks.md b/Benchmarks.md
new file mode 100644
index 0000000..7ee9812
--- /dev/null
+++ b/Benchmarks.md
@@ -0,0 +1,90 @@
+# Benchmarks
+
+Inference latency and throughput for winnow's two embedding models, measured on real hardware. Run with the `scripts/benchmark.py` script (included in the repo).
+
+---
+
+## Hardware
+
+| | |
+| :-- | :-- |
+| **CPU** | AMD / Intel (see per-run notes) |
+| **GPU** | NVIDIA GeForce RTX 2070 SUPER |
+| **Driver** | 570.172.08 |
+| **Container CUDA** | 12.8.1 |
+| **Host OS** | Debian 13 (TrueNAS LXC) |
+
+---
+
+## InsightFace Buffalo_L — face mode
+
+Detection (RetinaFace det_10g) + ArcFace embedding (w600k_r50) on a 640×640 image. This is the pipeline winnow runs for every face crop it evaluates.
+
+### RTX 2070 SUPER
+
+| Mode | Model load | Median latency | Throughput |
+| :-- | --: | --: | --: |
+| GPU (`FORCE_CPU=false`) | — | — | — |
+| CPU (`FORCE_CPU=true`) | 4.3 s | 102 ms | 9.8 img/s |
+
+*GPU results pending CUDA 12.8.1 image build.*
+
+### Notes
+
+- Latency measured over 30 runs after 5 warmup iterations.
+- Input: 640×640 synthetic image. The detection network processes the full input regardless of whether faces are found; timing is representative of real-world single-image throughput.
+- 320×320 input: CPU median 101 ms — detection runtime is dominated by the fixed model overhead, not image size at these resolutions.
+
+---
+
+## SigLIP ViT-B/16 — object mode
+
+`google/siglip-base-patch16-224` — 224×224 Vision Transformer used for object-mode diversity selection. Supports batched inference; GPU benefit scales with batch size.
+
+### RTX 2070 SUPER
+
+#### GPU (`FORCE_CPU=false`)
+
+| Batch | ms/batch | ms/img | img/s | p95/img |
+| --: | --: | --: | --: | --: |
+| 1 | — | — | — | — |
+| 4 | — | — | — | — |
+| 8 | — | — | — | — |
+| 16 | — | — | — | — |
+| 32 | — | — | — | — |
+
+*GPU results pending.*
+
+#### CPU (`FORCE_CPU=true`)
+
+| Batch | ms/batch | ms/img | img/s | p95/img |
+| --: | --: | --: | --: | --: |
+| 1 | 253 | 253 | 4.0 | 315 |
+| 4 | 821 | 205 | 4.9 | 214 |
+| 8 | 1566 | 196 | 5.1 | 214 |
+| 16 | 3181 | 199 | 5.0 | 205 |
+| 32 | 6175 | 193 | 5.2 | 199 |
+
+Model load: **18.8 s** (CPU; first load, no cache)
+
+CPU batching saturates quickly — throughput barely improves past batch 4 (~5 img/s ceiling). On GPU, large batches are expected to see significant throughput gains.
+
+---
+
+## Running the benchmark
+
+```bash
+# Inside the container — GPU mode:
+docker exec winnow python /app/scripts/benchmark.py
+
+# CPU-only mode:
+docker exec -e FORCE_CPU=true winnow python /app/scripts/benchmark.py
+
+# Or directly with docker run:
+docker run --rm --gpus all \
+  --entrypoint /app/.venv/bin/python \
+  -v /your/models:/insightface \
+  -e INSIGHTFACE_HOME=/insightface \
+  ghcr.io/sudolulo/winnow:latest \
+  /app/scripts/benchmark.py
+```
diff --git a/Home.md b/Home.md
index 2e2380e..4652f93 100644
--- a/Home.md
+++ b/Home.md
@@ -9,6 +9,7 @@ It runs fully headless in Docker, is configured entirely through environment var
 ## Pages
 
 - [[Setup]] — installation, Docker Compose configuration, GPU passthrough, environment variables
+- [[Benchmarks]] — GPU vs CPU inference latency and throughput (InsightFace, SigLIP)
 - [[Troubleshooting]] — common failures and fixes
 - [[FAQ]] — frequently asked questions
 
diff --git a/Setup.md b/Setup.md
index b7197b5..83ab2c8 100644
--- a/Setup.md
+++ b/Setup.md
@@ -25,7 +25,7 @@ This is the base URL of your Frigate instance, e.g. `http://192.168.1.10:5000`.
 
 | Tag | Arch | Acceleration |
 | :-- | :-- | :-- |
-| `:latest` | amd64 + arm64 | NVIDIA CUDA 13.3 (amd64) · requires [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) |
+| `:latest` | amd64 + arm64 | NVIDIA CUDA 12.8 (amd64) · requires [NVIDIA Container Toolkit](https://docs.nvidia.com/datacenter/cloud-native/container-toolkit/install-guide.html) · minimum driver 570 |
 | `:rocm` | amd64 | AMD ROCm · pass `/dev/kfd` + `/dev/dri` |
 | `:intel` | amd64 | Intel Arc / iGPU via OpenVINO · pass `/dev/dri`, set `OPENVINO_DEVICE=GPU` |
 | `:cpu` | amd64 + arm64 | CPU only · ~2 GB smaller · no GPU required |
@@ -151,7 +151,11 @@ environment:
   - OPENVINO_DEVICE=GPU   # omit to run OpenVINO inference on CPU (default)
 ```
 
-Set `OPENVINO_DEVICE=GPU` to target the Intel GPU. Omitting it (or setting `CPU`) runs OpenVINO on CPU — useful if you want Intel's optimised runtime without GPU passthrough.
+`OPENVINO_DEVICE` selects the compute device within the OpenVINO runtime:
+- `GPU` — Intel iGPU or Arc GPU (requires `/dev/dri` passthrough)
+- `CPU` *(default)* — CPU cores, via OpenVINO's CPU plugin which uses AVX512 and multi-threading. This is *not* the iGPU — it runs on the CPU itself, but OpenVINO's optimised kernels are meaningfully faster than standard CPUExecutionProvider.
+
+Use `CPU` if you want OpenVINO's faster CPU path without passing through the GPU device.
 
 ---
 
@@ -216,7 +220,7 @@ docker exec winnow cat /app/frigate_train/winnow.log
 | Variable | Default | Description |
 | :--- | :--- | :--- |
 | `FORCE_CPU` | `false` | Disable GPU — fall back to CPU for all inference |
-| `OPENVINO_DEVICE` | `CPU` | Intel variant only: set `GPU` to use Arc or iGPU |
+| `OPENVINO_DEVICE` | `CPU` | Intel variant only: `GPU` = iGPU/Arc (requires `/dev/dri`); `CPU` = OpenVINO CPU plugin (optimised CPU kernels, not the iGPU) |
 | `ENABLE_CACHE` | `true` | Cache computed embeddings to disk (speeds up re-runs on the same library) |
 | `CACHE_DIR` | `.if_cache` | Path for embedding cache and upload tracker files |
 | `HF_HOME` | *(system)* | HuggingFace model cache path (SigLIP) |
diff --git a/Troubleshooting.md b/Troubleshooting.md
index 74b83ec..ef4d723 100644
--- a/Troubleshooting.md
+++ b/Troubleshooting.md
@@ -118,7 +118,7 @@ Everything works on CPU but embedding computation is slower — typically tens o
      - render
    ```
 
-2. **Ensure `OPENVINO_DEVICE=GPU` is set** — without it, OpenVINO defaults to CPU even with device passthrough.
+2. **Ensure `OPENVINO_DEVICE=GPU` is set** — without it, OpenVINO runs on the CPU (via its optimised CPU plugin), not the iGPU, even with `/dev/dri` passed through.
 
 3. **Verify OpenVINO sees the device**:
    ```bash