Electronic Engineer's Notebook

Floating Point Processor Performance


Floating Point DSP Trend

Although the market for fixed point DSP processors is much larger, floating point devices are still key for many applications. This page attempts to show how their processing power has progressed over time.

One days time, raw floating point performance was only provided by specialist DSP devices. These days most processors include dedicated floating point units. The latest RISC processors provide a performance at least comparible with specialist devices if not exceeding them.

MFLOPS vs first year of production

This shows how the peak million floating point operations per second (MFLOPS) has improved over the years. The horizontal scale is approximately year of production. Peak MFLOPS is, of course, a very crude measure of processor capability. It is simply the sum of the peak processing power of all of the floating point units. The sustained performance that can be achieved will depend on to the extent that the floating point units can be employed. That is whether data can be got on and off chip fast enough, whether there is enough internal bus bandwidth to shift the data between the different function blocks, whether there is enough internal memory and registers to store intermediate results, whether instructions can be issued fast enough to control the processing. etc. It also depends on the mix of operations required by the application. Another factor to take into account, if the programmer is not working at the assembly level, is the efficiency of the compiler that is employed.

Below is a table giving a bit more information on the progression of the technology and its capabilities.

Year

Device

Peak MFLOPS

1K Cmplx FFT

Technology

1985

68020

0.2

1.5 mm

1986

T800 Transputer

3

1.5 mm

1987

68030

0.5

1.2 mm 270,000 transistors

1989

68040

5.6

0.8 mm 1.2 million transistors

1989

TMS320C30 33MHz

33

3.04ms

1 mm

1990

Intel i860

80

745 ms

1 mm, 1 million transistors

1992

Intel i860XP

100

596 ms

0.8 mm

1995

PowerPC 603e

200

840 ms

0.5 mm 2.6 million transistors

1996

21060 Sharc

80 (120 FFT)

460 ms

1999

21065L Sharc

120 (180 FFT)

310 ms

1999

TMS320C6701 167MHz

1000

108.33 ms

0.18 mm

2000

TMS320C6711

900

120.4 ms

0.18 mm

2000

PowerPC 7400 375MHz

3000

25 ms

0.18 mm

2001

21160 Hammerhead

600

92ms

2001

TMS320C6712

600

180.6ms

0.18 mm

2001

ADSP-TS001 TigerSharc

900

69 ms

0.18 mm


Top

Home


Last updated 22 October, 2000

© 2000 Michael Wells