InferEdge
Inference as a bounded transaction, not a kernel call
AI inference accelerator for energy-bound sites. Each request is admitted against a joule budget and completes with proof or aborts explicitly.
Interposer stack
- Inside the contract
- Beyond the bound
- Light stops at the bar
Read the architecture
Each state is entered explicitly, failure is a defined transition rather than an absence, and what is drawn is specified rather than measured.
- Ordered step
- Only exit
- Measured epoch
- Flagged for re-measure
- Light = the digest being extended
What holds, and why
-
Energy contract admission
Budget, model tier, clock, and checkpoint policy agreed before work starts
-
Completion state machine
Every request occupies an explicit state; failure is a defined transition
-
Checkpointed recovery
Restore validates integrity and contract before resume
-
Proof-carrying output
Result digest bound to authorised execution context
Specification
Pre-silicon. Architecture specified; RTL in progress; FPGA next. Figures are design targets, not measured silicon results.
| Status | RTL in progress |
|---|---|
| Power envelope | Sub-15 W (design target) |
| Numeric formats | INT8 / INT4 (design target) |
| Execution model | Completion transaction under energy contract |
| Recovery | Checkpoint and validated restore |
| Output | Result plus attestation bundle |
| Validation route | FPGA prototype before ASIC |
- Brownout-tolerant sites
- Energy-constrained edge
- Proof-carrying inference
- Offline-capable execution
- Narrowband telemetry sites
Talk to the design team
The specification is still open to review.