Analog dynamics.
Digital coordination.
A new route to NP-hard optimization.
AATLAS brings together reconfigurable analog circuits and digital coordination to investigate energy-efficient computing for hard optimization problems.
Lower states represent better candidate solutions.E(s) → minimum
Let the circuit
search for a solution.
Scheduling tasks, partitioning a graph or finding a short route can become difficult as a problem grows. AATLAS investigates how the dynamics of an electronic network can help search for good solutions.
The idea is to encode a problem in interacting analog elements, let the network evolve, and use digital hardware to configure, coordinate and read out the computation.
Read the ANR project summaryFPAA Field-programmable analog array
Reconfigurable circuits implement Hopfield and Ising networks. Their continuous dynamics perform the core optimization process.
FPGA Field-programmable gate array
Digital logic provides control, interfaces and pre- and post-processing, connecting analog cores within a larger system.
From a single core
to a hybrid solver.
Three connected work packages, from analog network dynamics to system-level evaluation.
Computing with
analog dynamics
WP1 · Georgia Tech–CNRSImplement Hopfield and Ising networks on a single FPAA and investigate strategies for exploring the optimization landscape.
Connecting
multiple cores
WP2 · Institut Jean LamourDevelop a hybrid FPAA–FPGA architecture that links smaller computing blocks to address larger optimization problems.
Understanding
the complete system
WP3 · CentraleSupélecStudy nonlinear dynamics and hardware effects, map optimization problems onto the architecture, and evaluate solution quality, convergence time and energy consumption.
Quality. Time. Energy.
Three criteria for evaluating a hybrid solver.
How close is the result to the best available solution?
How long does the complete computation take?
How much electrical energy does it require?
A hard problem.
A simple example.
Max-Cut asks how to split a network into two groups so that as many connections as possible run between them.
Select a node to change its group.
Make more connections
cross the divide.
Each node has two possible states. Changing one can improve several connections while making others worse.
between the two groups
Educational example computed in your browser, not an FPAA measurement. AATLAS studies how analog dynamics can search for good solutions; reaching the global optimum is not guaranteed.
The people behind
the computing.
Three partner groups bring together analog circuit design, reconfigurable electronics and nonlinear dynamics.
Georgia Tech–CNRS
& Georgia Tech
IRL 2958, Metz · School of ECE, Atlanta
FPAA computing cores, Hopfield and Ising networks, and experimental nonlinear dynamics.
- Alexandre LocquetProject coordinator · CNRS Research Scientist
- Jennifer HaslerProfessor · Georgia Tech
- David CitrinProfessor · Georgia Tech
- Muhammad UsmanPhD researcher
- Harikrishna KayithaPhD researcher
- Corentin BounonPhD researcher
- Olivier KonneCNRS technician · Circuit design and fabrication
CentraleSupélec
LMOPS
Laboratoire Matériaux Optiques, Photonique et Systèmes · Metz
Modelling, nonlinear networks and evaluation of the hybrid optimization system.
- Damien RontaniProfessor · Partner lead
- Nicolas MarsalProfessor
- Eric Oliveira GomesPhD researcher
Institut Jean Lamour
Université de Lorraine–CNRS · Nancy
FPGA architectures, digital coordination and the integration of multiple analog computing cores.
- Slaviša JovanovićProfessor · Partner lead
- Hassan RabahProfessor
- Manuel DjekoulaPhD researcher
Let’s discuss the research.
Scientific enquiries and collaboration · Alexandre Locquet, project coordinator