Tuesday November 22nd, 2016
Location: Panther Hollow Conference Room, CIC - 4th Floor
As FPGAs have grown in size and capacity, FPGA memory systems have become both richer and more diverse in order to support the increased computational capacity of FPGA fabrics. Gone are the days of a building BRAM caches backed by a single DRAM: modern FPGA systems feature multiple off-chip memory resources, often with asymmetric characteristics, and new fabric resources like eSRAMs. Using complex FPGA memory resources, and using them well, has become commensurately more difficult, especially in the context of legacy designs ported from smaller, simpler FPGA systems. This growing complexity necessitates the resource-aware compilers that can make good use of memory resources on behalf of the programmer. In this work, we introduce the LEAP memory compiler, which can synthesize application-optimized cache networks for systems with multiple memory resources, enabling user programs to automatically take advantage of the expanded memory capabilities of modern FPGA systems. In our experiments, the optimized cache network achieves up to 46% performance gain for throughput-oriented applications and 15% performance gain for latency-oriented applications, while increasing design area by less than 5%.
This work is joint with Hsin-jung Yang and Michael Adler.
Kermin Fleming (BS, BS, MS, 2006) obtained his Ph. D. in EECS from MIT in 2012, with a focus on network synthesis in FPGA systems. He now works are Intel, where he researches reconfigurable architectures and software systems.