A. H. C. Barros

Implementation of a double-precision multiplier accumulator with exception treatment to a dense matrix multiplier module in FPGA

Recently, the manufactures of supercomputers have made use of FPGAs to accelerate scientific applications [16][17]. Traditionally, the FPGAs were used only on non-scientific applications. The main reasons for this fact are: the floating-point computation complexity; the FPGA logic cells are not sufficient for the scientific cores implementation; the cores complexity prevents them to operate on high frequencies. Nowadays, the increase of specialized blocks availability in complex operations, as sum and multiplier blocks, implemented directly in FPGA and, the increase of internal RAM blocks (BRAMs) have made possible high performance systems that use FPGA as a processing element for scientific computation [2]. These devices are used as co-processors that execute intensive computation. The emphasis of these architectures is the exploration of parallelism present on scientific computation operations and data reuse. In major of these applications, the scientific computation uses, in general, operations of big floating-point dense matrices, which are normally operated by MACs. In this work, we describe the architecture of an accumulative multiplier (MAC) in double precision floating-point, according to IEEE-754 standard and we propose the architecture of a multiplier of matrices that uses developed instances of the MACs and explores the reuse of data through the use of the BRAMs (Blocks of RAM internal to the FPGAs) of a Xilinx Virtex 4 LX200 FPGA. The synthesis results showed that the implemented MAC could reach a performance of 4GFLOPs.

High Performance Implementation of RTM Seismic Modeling on FPGAs: Architecture, Arithmetic and Power Issues

This work presents a case study in the oil and gas industry, namely the FPGA implementation of the 2D reverse timing migration (RTM) seismic modeling algorithm. These devices have been largely used as accelerators in scientific computing applications that require massive data processing, large parallel machines, huge memory bandwidth and power. The RTM algorithm enables you to directly solve the acoustic and elastic waves problems with precision in complex geological structures, demanding a high computational power. To face such challenges we suggest strategies such as reduced arithmetic precision, based on fixed-point numbers, and a highly parallel architecture are suggested. The effects of such reduced precision for storage/processing data are analyzed in this chapter through signal-noise ratio (SRN) and universal image quality index (UIQI) metrics. The results show that SRN higher than 50dB can be considered acceptable for a migrated image with 15 bits word size. A special stream-processing architecture aiming to implement the best possible data reuse for the algorithm is also presented. It was implemented by an FIFO-based cache in the internal memory of the FPGA. A temporal pipeline structure has also been developed, allowing that multiple time steps to be performed at the same time. The main advantage of this approach is the ability to keep the same memory bandwidth needs of processing just one time step. The number of time steps processed at the same time is limited by the amount of FPGA internal memory and logic blocks. The algorithm was implemented on an Altera Stratix 260E, with 16 processing elements (PEs). The FPGA was 29 times faster than the CPU and only 13% slower than the GPGPU. In terms of power consumption, the CPU+FPGA was 1.7 times more efficient than the GPGPU system.

Performance Evaluation Model based on Precision Reduction and FPGAs Applied to Seismic Modeling

The recent increase in computing power of FPGAs has allowed its use in areas such as seismic data processing. Additionally, besides the capability of performing computations in parallel way, FPGAs also support application-specific number representations. In this type of application, in order to achieve better performance, instead of using the floating-point standard, usually the processing and storage of data is done using the fixed point standard. However, the change of representation can cause a degradation in the quality of the results. In the petroleum industry, a seismic image of poor quality can represent an erroneous interpretation of the subsurface, resulting in catastrophic losses. For this reason, it is essential that the quality of data obtained from the seismic data processing for low precision can be evaluated within reliable technical criteria. In this paper, a real case study was used in order to evaluate the efficiency of two different metrics applied to this seismic application based on RTM algorithm. The main strategy is to explore the precision reduction in terms of SNR (Signal-to-Noise Ratio) and UIQI (Universal Image Quality Index) metrics, in order to improve the performance of the system. Results show a performance gain of 50% compared with the architecture implemented in hardware using floating point standart IEE754.

Poster: high performance FPGA-based implementation of the seismic modeling of the RTM algorithm

Hardware accelerators like GPGPUs and FPGAs have been used as an alternative for the conventional computing architectures (CPUs) in scientific computing applications and have shown considerable speed-ups. In this context, this poster presents a solution that takes advantage from FPGAs flexibility to explore efficiently data reuse, parallelization in both time and space domains for the first processing stage of the RTM (Reverse Time Migration) algorithm, the seismic modeling. In order to obtain a benchmark for our FPGA implementation, we also implemented the same algorithms for a CPU and GPGPU architecture. Our results showed that the FPGAs are a feasible platform for this set of applications. The experimental results have shown a 1,67x speed-up when compared to a Tesla C1060 GPGPU and a 25,79x speed-up when compared to an AMD Athlon 64 X2 CPU.