Hartley Ultrafast
Hartley builds photonic neural network computers that perform AI inference in roughly 160 nanoseconds, enabling decision-making at speeds previously unattainable with conventional digital hardware. Its Babbage proof-of-concept system couples optoelectronic neural processing units with Altera's FPGA platform, targeting defence, industrial control, and capital markets applications where latency is mission-critical.
- Artificial Intelligence
- Defense Tech
- Hardware
Funding
Founders
Founder details are not available yet.
Product
Problem
Conventional AI inference on digital hardware—even the latest GPUs—introduces microsecond-to-millisecond latencies that are too slow for real-time, mission-critical decision-making in domains such as defence, industrial control, and high-frequency trading. This latency gap prevents AI from being embedded in systems where speed and determinism are non-negotiable.
Solution
Hartley delivers photonic neural network computers that perform AI inference in roughly 160 nanoseconds, bringing decision-making into a new domain of speed. Its Babbage platform combines optoelectronic neural processing units (ONPUs) with Altera's FPGA technology, enabling hybrid neuromorphic-plus-digital computing that is both ultrafast and programmable. The system uses in-memory compute, where model weights live directly in the physical fabric of the machine, eliminating the need to shuttle data between memory and processor. Hartley also provides a digital twin training workflow, allowing customers to train models offline using standard machine learning toolchains and then upload weights directly to the hardware for inference without retraining.
Target Audience
Primary customers are defence, industrial control, network intelligence, and capital markets organizations that require real-time, deterministic AI decision-making at nanosecond timescales, as well as system integrators and developers working within existing FPGA ecosystems.
Features
- Optoelectronic neural processing units (ONPUs) that combine photonics and electronics for nanosecond-scale inference latency
- Babbage proof-of-concept system demonstrating 176-nanosecond inference on a 4-neuron-wide, 3-layer-deep neural network, faster than the latest GPU hardware
- Clockwork time-synchronisation system that maintains deterministic timing across optical and electronic domains
- Ultrafast analog electronic Decision Engine that returns ordinal data with extremely low latency
- Digital twin training workflow that enables all-digital, offline model development using standard ML tooling, with direct weight upload to hardware
- Integration with Altera's FPGA platform, providing a drop-in extension for existing FPGA deployments with familiar development flows