Floating-point emulation: faster than hardware?

 
Last year, I was at the decision point. I needed to select a processor for the next version of Critical Link's MityDSP "custom-off-the-shelf" CPU platform—basically a collection of integrated building blocks (DSP, FPGA, FLASH, SDRAM memory subsystems, etc.), with configurable I/O that we adjust to fit individual customer requirements.

The original MityDSP uses TI's TMS320C6711 floating point processor. This served—and still serves—many of our customers' needs. But some customer requirements began to necessitate more powerful board options. We boosted our original Mity design with a larger FPGA (Xilinx XC3CS1000 vs. the XC3CS400) and increased the CPU's RAM and flash. But for some of our more processing-intensive applications, we were running up against the original MityDSP's overall limits. One of the major limiting factors was the C6711 family's maximum clock rate (200 MHz at the time—less than 1/3 the max clock rate of TI's C64x family).

We knew we needed a new set of building blocks—what would become the MityDSP Pro—to support faster processing speeds. The question was whether to stick with a floating-point processor, or move to a fixed-point DSP and handle floating-point operations through emulation.

The processor we were most interested in was the TI C6454/55. One of the primary attractions was the high clock rate—over 1 GHz with the C6455. Another was the C64x's significantly improved dual memory busses. We were also strongly motivated by our long experience with the C6711. The similarity in architecture and peripherals between the TI C64x and the TI C6711 were obviously going to provide us an advantage.

Click here to read more ...