Trusted Compute Platform
Four layers that make an inference result verifiable
Root of trust, inference engine, infrastructure SoC, and fleet verification plane, specified together so a field result can be checked against the hardware, firmware, and model that produced it.
Advanced package
- Trust boundary
- Supporting logic
- Signed data in flight
Read the architecture
Each layer depends only on the guarantees beneath it, and what is drawn is specified rather than measured.
-
Attested Infrastructure Inference
Fleet verification before workflow admission · Platform
-
InferEdge
Joule-bounded inference as a completion transaction · Accelerator
-
SecureGrid
Metrology and inference in one trust domain · SoC
-
TrustCore
Per-die identity, measured boot, result signing · Root of trust
- Current stage
- Reject path
- Light = hand-off and measurement
What holds, and why
-
Silicon-derived identity
PUF identity from manufacturing variation, no stored key to extract
-
Measure before execute
Boot, firmware, and model artefacts verified before inference runs
-
Energy-budgeted work
Requests admitted only against an accepted joule budget and model tier
-
Proof-carrying results
Each export binds result digest to device, firmware, and model state
Specification
Pre-silicon. Architecture specified; RTL in progress; FPGA next. Figures are design targets, not measured silicon results.
| Platform status | Architecture defined; RTL in progress |
|---|---|
| Target inference envelope | Sub-15 W (design target) |
| Numeric formats | INT8 / INT4 (design target) |
| Identity | PUF-derived, per-die |
| Cryptography | Post-quantum and classical suites |
| Target service life | 10-20 years in field |
| First silicon | Not before 2029 |
- Revenue metering
- Distribution automation
- Industrial monitoring
- Remote radio sites
- Constrained-edge inference
Talk to the design team
The specification is still open to review.