AM-FL · Autonomous Driving
In a moving vehicle,
a late decision is a dangerous one.
Deterministic neural perception and decision for autonomous driving — fixed latency, no OS on
the decision path, low power. A small, auditable engine built for a functional-safety process.
Powered by the FLVH inference engine.
Why variable latency is a safety problem
Autonomous perception runs in a vehicle that does not wait. A decision that arrives at a known, bounded time is safe; one that depends on an OS scheduler, a framework and system load is not — and it is far harder to argue before a safety authority.
A GPU/SoC stack running Linux or a full automotive middleware brings high power, variable latency, and a software base measured in millions of lines — an enormous surface to secure and to certify. AM-FL removes it from the decision path: no OS, fixed latency, low power, auditable from the source. And in an electric vehicle, every watt counts as range — under 1 W of perception draws almost nothing from the battery.
Why AM-FL fits functional safety
Bounded, deterministic latency — the decision arrives in a known time, every time; the basis of a defensible safety argument. No OS, no driver, no framework on the decision path — nothing to crash, drift, or exploit, and a far smaller safety case. Auditable engine — under 5,000 lines of Verilog a certification team can read by hand, supporting an ISO 26262 / ASIL process. Model update by binary swap — push a new DDR binary, no FPGA re-synthesis.
AM-FL vs. SoC + OS perception
Criterion
SoC + OS + framework
AM-FL
Decision latency
Variable (scheduler, load)
Fixed by construction
Jitter
Not guaranteed
±12 µs measured
Software on decision path
OS + drivers + middleware
None
Code to certify
Millions of lines
<5,000 lines Verilog
Functional safety (ISO 26262)
Heavy
Architecture-ready
Where AM-FL runs
◉
Perception at Fixed Latency
Object and pedestrian detection that returns a result in a known, bounded time — the property a safety case needs, regardless of scene complexity or system load.
⏱
Emergency Decision
Time-critical decisions (brake / steer-assist trigger) computed in <1 ms, on a path with no OS to preempt it. The reaction window is deterministic.
⛁
Sensor Front-End
On-board preprocessing and feature extraction from camera/radar before the main compute — filter and reduce at the edge, at fixed latency and under one watt.
◐
In-Cabin Monitoring
Driver attention / occupancy inference running locally and continuously — no cloud, no video off the vehicle (privacy by design), no OS update that can disable it.
The engine is proven on silicon
AM-FL runs the same FLVH inference core validated bit-exact on physical hardware — Zynq-7020 and AMD Kria KV260 (ZU5EV). Reference benchmarks (MNIST): 98.05% MLP, 99.06% CNN, 0.00% delta vs float64; jitter ±12 µs over 10,000 runs (Zynq-7020). Your perception model is trained for your sensors and your task — the engine, the determinism and the power envelope are already qualified.
Note: published benchmarks are reference workloads. ISO 26262 / ASIL assessment is addressed per programme; the architecture is designed to support such a process, it is not in itself a certification.
Evaluate AM-FL on your bench
A sealed, signed AM-FL product on your KV260 — measure the latency, jitter and bit-exactness on your own hardware before integration.
Evaluation license: $10,000 · 12-month KV260 bench.
Production: per-vehicle royalty or programme license — by discussion.
Request an evaluation →