In the first year of the project, the relevant technical outputs were produced in the frame of Workpackage 2, which develops a custom intelligence block to be integrated into the sensor node. It will contain a custom-designed accelerator for energy-efficient probabilistic learning, controlled by a RISC-V microcontroller.
In the second year the has seen the project progress in all SUSTAIN technologies. The publishing of results has been active, as can be seen in the list of publications. Also, a high number of 14 Deliverables have been produced during the second year.
We have developed a strategy towards a system model of the distributed node to be developed and which will integrate the intelligence of the overall system (D2.1). Particularly, the strategy comprises three phases to a) build an initial framework using commercial sensor nodes and a first meta-intelligence, b) to customize and improve the framework with intelligence mechanisms developed in SUSTAIN and c) to finally build the framework with final nodes and to test the system in-situ.
As processor for the system, the A-core processor developed by Aalto University will be utilized. Node-level intelligence is planned to be implemented using probabilistic circuits. For the testing of the node-level intelligence and meta-learning, a simulator developed at Trento University will be employed.
A first proposal for probabilistic node intelligence has been made in D2.2. Particularly, Our objective was to evaluate how to increase the efficiency of probabilistic inference on hardware, to facilitate the implementation of PCs, using approximate computing blocks that are efficient on hardware. A path towards more efficient PC inference on hardware is to perform inference as in software, i.e. alternating between logarithm and linear computations. To achieve this, we propose to build an approximate computing framework dedicated to PCs, leveraging Addition As Int (AAI). Generally, all multipliers can be replaced with AAI to save energy. However, that may have a dramatic impact on the accuracy of the model, because certain nodes require a very high resolution to be computed. Instead, we target the development of a dedicated methodology to use AAI approximate multipliers in an optimal way for efficient PC inference.
We evaluated if we could reduce the number of bits of PC inference, and if replacing all multipliers by AAI approximate versions would have a large impact on the model’s accuracy. We took four benchmarks as an example (NLTCS, Jester, DNA, Book), among the most used benchmarks used in the literature for probabilistic models. In some cases, it is not possible to tolerate error in the model. Instead, we would like to be able to safely replace part of the multipliers in the PC while having no or a very limited impact on accuracy. That is why in the second experiment, we propose and evaluate an error compensation technique and dedicated replacement methodology, to safely replace multipliers by AAI also in the case of MAR query.
In parallel of the development around approximate computing for PCs, Aalto University has successfully tested their first open-source RISC-V processor named A-core. The silicon chip has been taped-out in a 22nm technology in December 2022 and fully tested during the summer of 2023 (publications are ongoing). This processor can serve as a baseline for future integration withing the sustAIn project.
