The Defense Sciences Office at the Defense Advanced Research Projects Agency (DARPA) is soliciting research proposals in technologies for the acceleration of scientific simulations of physical systems characterized by coupled partial differential equations (PDEs). The Accelerated Computation for Efficient Scientific Simulation (ACCESS) Program seeks innovative ideas for computational architectures that will achieve the equivalent of petaflops performance in a benchtop form-factor and be capable of what traditional architectures would define as “strong” scaling for predictive scientific simulations of interest. DARPA expects achieving these goals will require the parallel development of non-traditional component technologies exploiting novel hybrid analog/digital techniques, algorithms, instruction sets, controllers, and the integration and optimization of these components within prototype systems. Specifically excluded is research that primarily results in evolutionary improvements to the existing state of practice.
Many problems in plasma physics and fluid systems are governed by long-range forces, local interactions, and non-linear phenomena that span many dimensions and timescales and require multiple coupled non-linear PDEs to be adequately modeled. Extreme examples include high energy density z-pinch1,2 and tokamak3,4 plasma systems that require fully kinetic models to resolve the desired physics over time and spatial scales spanning orders of magnitude. The ultimate goal of the ACCESS program is to demonstrate new specialized benchtop technology that can solve large problems in these kinds of complex physical systems on the hour timescale, compared to normal computational methods that require full cluster-scale supercomputing resources on the weeks-to-months timescale. The demonstration of such novel prototypes may provide foundational technologies for specialized scientific computing systems beyond Moore’s law and could transform how scientific simulations are used for both design and discovery of complex physical systems.
The design and development of the desired prototypes are envisioned to leverage advances in optics, MEMS, additive manufacturing, and other emerging technologies to develop new nontraditional analog and digital computational means and to overcome some of the current known limitations of these means, such as precision and stability. Of particular interest are hybrid analog/digital architectures that replace numerical methods and memory-intensive computational parallelization with nonlinear and/or intrinsically parallel physical processes to perform computations. Other novel approaches are also of interest to DARPA if they can achieve the desired capability. Specifically excluded are black-box solvers which require prohibitively large sets of training data and fail to generalize well to outside parameter regimes.
It is expected that there will be a tradeoff between the generalizability of the computational means and the acceleration expected versus classical computers. For example, an approach that computes using direct physical analogy to the system of interest may be unsuitable for any other application. On the other hand, a digital co-processor that accelerates the calculation of exponential functions will have wide applicability for scientific computation but may not meaningfully accelerate communications-limited simulations. Proposals should explain these sensitivities in detail within the context of the chosen benchmarking problems and their physical systems.
To assess the technical potential for these approaches to achieve equivalent “petaflops-onbenchtop” performance, the ACCESS program is focused on enabling technologies for the demonstration of end-to-end simulations of proposer-defined benchmarking problems within plasma and fluid physical systems. Proposers must choose a physical system to model and apply their technical approach to at least one benchmarking problem. Candidate benchmarking problems in these physical systems are discussed further below and in detail in Appendix 1. ACCESS program results could lead to follow-on efforts culminating in an integrated end-to-end prototype system.