Souped-up FPGAs execute FFTs

Altera's cores comply with the IEEE 754 standard for binary floating-point arithmetic, said Martin Langhammer, chief scientist and driver for DSP development at Altera (San Jose, Calif.). The Stratix family of devices, architected with up to 115,000 logic cells and 7 Mbits of memory, are particularly good for operations with a big mantissa, a numerical value whose floating decimal point can make it extraordinarily precise, he said.

While the IEEE 754 single-precision format calls for 1 sign bit, 8 exponent bits and 23 mantissa bits for both data and "twiddle" manipulation, Altera's floating-point cores use the standard's extended precision arithmetic, with 1 sign bit, 8 exponent bits and 31 mantissa bits. Thus, the Altera FFT floating-point core will perform 36 x 36-bit-wide multiplies at 250-MHz cycle rates. Altera's Radix 4 core will achieve a 1,024-point transform in 25 microseconds, while the Radix 2 will take 50 microseconds.

The cores are meant to support designs that require powerful FFT processing capabilities, such as military radar, signal intelligence and industrial and medical systems. "Our customers previously used either off-the-shelf standard products or multiple compute engines-computer farms-with quad Altivec systems [PowerPC processors with vector-processing extensions] to process our wideband data," Langhammer said.

Satellite imaging is one of the biggest applications for floating-point processors , said Will Strauss, principal analyst at Forward Concepts (Tempe, Ariz.), a DSP market tracker. FFTs and FIR filters are the two most popular DSP operations, Strauss said. High resolution-a 65,000-point FFT-is required to find an airplane or missile battery hidden among the "ground clutter" on a battlefield, Strauss said.

Dedicated FFT devices like the Butterfly processor marketed by Sharp Microelectronics (Vancouver, British Columbia) are not "cheap chips," Strauss said. "They cost a few hundred dollars each." A few custom parts use an FPGA architecture to execute a floating-point operation, Strauss said, like one marketed by Texas Memory Systems Inc. (Houston), but Altera is the first to make such a part a "catalog item."

Altera said its Radix 4 core uses only about 10 percent of the resources in a midrange Stratix FPGA. The Radix 2 is half the size of the Radix 4. Additional versions of the cores are under development, the company said.