IBM & AKD1000, page-20

Not to forget Airbus and the Neuravis ESA sponsored AKIDA project:

Evaluation of Neuromorphic computing
technologies for very low power AI/ML applications

Roland Brochard
Jérémy Lebreton
Lucas Marti
Nicolas Menga
Airbus Defence & Space
Toulouse, France
Harvey Gomez
Airbus Defence & Space
Ottobrunn, Germany

Gilles Bezard
Alf Kuchenbuch
Douglas McLelland
BrainChip
Toulouse, France
Kenneth Ostberg
David Juhasz
Frontgrade Gaisler
Göteborg, Sweden

Nicolas Bourdis
Florian Corgnou
Gregor Lenz
Karl Vetter
Neurobus
Paris, France
Laurent Hili
European Space Agency
Noordwijk, Netherlands

Abstract—By mimicking the human brain, neuromorphic
computing offers an energy efficient alternative to traditional
Von Neumann computer architectures. The objective of this
project is to evaluate the potential of the technology using neural
network models optimized for BrainChip's Akida. Three image
processing use cases are addressed: dense optical flow for
planetary landing, monitoring of a satellite model-based
tracking algorithm, and ship detection by Earth Observation
satellites. We assess their implementation on both the Akida
1500 board, and in an embedded Intellectual Property (IP)
version ported on a radiation-hardened FPGA demonstrator.
To that goal a generic hardware-software interface is ported on
the target. Benchmarks are conducted assessing key
performance metrics including accuracy, processing time and
latency....

...GIGAkida, a backend for Akida was implemented to perform
benchmarks and identify limitations. The Akida engine API
differs from GIGA’s layer-by-layer function calls, requiring
binary blobs to be generated offline using the Python
toolchain. GIGAkida merges layers to minimize data
transfers between Akida and RAM (Random Access
Memory), and handles large tensors by splitting them into
smaller tiles. The workflow includes an optional offline
processing pass to generate binary blobs, ensuring backward
compatibility with previous projects. The GIGAkida backend
overcomes Akida v1 limitations by splitting large tensors,
merging layers, and providing software fallbacks for
unsupported layers...

VII. CONCLUSION

We successfully tested three different kinds of neural
networks on Akida v1 accelerator. The acceleration factor
with respect to a regular CPU is interesting for highly
demanding computer vision workloads. The Akida v1 is well
suited for classification tasks but imposes limitations for
regression tasks such as dense optical flow. We demonstrated
that we need at least 8bits quantization to address this class
of problems which means it will be fully possible with Akida
v2. Execution time for AI models is usually deterministic
with dense tensor processors, but since SNN rely on sparse
processing capabilities, it introduces a dependency to
processed data. In addition, it was not clear that SNN could
be wrapped into a dense, layer by layer, API like GIGA but
we demonstrated it is possible. This is important for
portability as the choice of a platform depends on many
factors, not only processing performance or efficiency.