Section: Research Program

New Models for New Technologies

Over the past 25 years there have been several hardware-architecture generations dedicated to massively parallel computing. We have contributed to them in the past, and shall continue doing so in the Dreampal project. The three generations, chronologically ordered, are:

• Supercomputers from the 80s and 90s, based on massively parallel architectures that are more or less distributed (from the Cray T3D or Connection Machine CM2 to GRID 5000). Computer scientists have proposed methods and tools for mapping sequential algorithms to those parallel architectures in order to extract maximum power from them. We have contributed in this area in the past.

• Parallelism pervades the chips! A new challenge appears: hardware/software co-design, in order to obtain performance gains by designing algorithms together with the parallel architectures of chips adapted to the algorithms. During the previous decade many studies, including ours in the Inria DaRT team, were dedicated to this type of co-design. DaRT has contributed to the development of the OMG MARTE standard (http://www.omgmarte.org) and to its implementation on several parallel platforms. Gaspard2, our implementation of this concept, was identified as one of the key software tools developed at Inria: http://www.inria.fr/en/centre/lille/research/platforms-and-flagship-software/flagship-software.

• The new challenge of the 2010s is, in our opinion, the integration of dynamic reconfiguration and massive parallelism. New circuits with high-density integration and supporting dynamic hardware reconfiguration have been proposed. In such architectures one can dynamically change the architecture while an algorithm is running on it. The Dynamic Partial Reconfiguration (DPR) feature offered by recent FPGA boards even allows, in theory, to generate optimized hardware at runtime, by adding, removing, and replacing components on a by-need basis. This integration of dynamic reconfiguration and massive parallelism induces a new degree of complexity, which we, as computer scientists, need to understand and deal with order to make possible the design of applications running on such architectures. This is the main challenge that we address in the Dreampal project. We note that we adress these problems as computer scientists; we do, however, collaborate with electronics specialists in order to benefit from their expertise in 3-D FPGAs.

Excerpt from the HiPEAC vision 2011/12

“The advent of 3D stacking enables higher levels of integration and reduced costs for off-chip communications. The overall complexity is managed due to the separation in different dies, independently designed.”

FPGAs (Field Programmable Gate Arrays) are configurable circuits that have emerged as a privileged target platform for intensive signal processing applications. FPGAs take advantage of the latest technological developments in circuits. For example, the Virtex7 from Xilinx offers a 28-nanometer integration, which is only one or two generations behind the latest general-purpose processors. 3D-Stacked Integrated Circuits (3D SICs) consist of two or more conventional 2D circuits stacked on the top of each other and built into the same IC. Recently, 3D SICs have been released by Xilinx for the Virtex 7 FPGA family. 3D integration will vastly increase the integration capabilities of FPGA circuits. The convergence of massive parallelism and dynamic reconfiguration in inevitable: we believe it is one of the main challenges in computing for the current decade.

By incorporating the configuration and/or data/program memory on the top of the FPGA fabric, with fast and numerous connections between memory and elementary logic blocks ($\sim$10000 connections between dies), it will be possible to obtain dynamically reconfigurable computing platforms with a very high reconfiguration rate. Such a rate was not possible before, due to the serial nature of the interface between the configuration memory and the FPGA fabric itself. The FPGA technology also enables massively parallel architectures due to the large number of programmable logic fabrics available on the chip. For instance, Xilinx demonstrated 3600 8-bit picoBlaze softcore processors running simultaneously on the Virtex-7 2000T FPGA. For specific applications, picoBlaze can be replaced by specialized hardware accelerators or other IPs (Intellectual Property) components. This opens the possibility of creating massively parallel IP-based machines.