Benchmarks
Numbers with their conditions, including the slow rows.
Your model runs encrypted on customer sites, edge devices and partner clouds. These are the costs in latency, accuracy and memory of keeping it encrypted there.
Scope: computer vision, classification and time-series models with encrypted weights, by model and hardware.
Benchmark configurationbenchmark.yaml
benchmark: model_file: "your-model.onnx" task: classification input_shape: [1, 3, 224, 224] batch_size: 1 target_hardware: "your-target-hardware" compare: - plaintext - encrypted report: - latency - accuracy_delta - peak_memory output: request_accessReview encrypted inference on your model and target hardware with our technical team.
Method
What every row will state
- Encryption scheme and parameters
- Quantisation applied and its accuracy effect
- How latency was measured, warm or cold, and over how many runs
- Exact hardware, not a hardware class
What these numbers mean for your cadence
- Per study
- In scope
- Per batch
- In scope
- Per event
- In scope
- Line-speed video
- Out of scope
Want your model on this table?
We tell you what it costs in latency and what it buys in protection.