Chapter 6: DSP And Its Impact On Technology Book: Processor Design Systems On Chip Computing For ASICs And FPGAs By Jari Nurmi Slides Prepared by: Omer Anjum
Introduction
The early beginning g of DSP DSP is relatively a new science Its roots go to the group of universities with the rediscovery of FFT in mid 1960s To process computationally intensive DSP kernels in real time was even not possible with main frame computers of that time
An Algorithm For Machine Calculation For Complex Fourier Series
Earlier computer to do research in DSP
Earlier with the advancement, microprocessors were the choice to do some simple DSP tasks in real time Mini computers and Array processors were state of the art for DSP research community Breakthrough that made the DSP possible was the combination of a microprocessor and a hardware multiplier With in last 30 years sharing of ideas between DSP research and circuit community and invention of transistor revolutionized the world with commercially available DSP
That discipline which has allowed us to replace a circuit previously composed of a capacitor and a resistor with two anti-aliasing filters, an A-to-D and a D-to-A converter, and a general purpose computer (or array processor) so long as the signal we are interested in does not vary too quickly. Dr. Tom Barnwell
The DSP revolution DSPs were primarily a university curiosity in the 1960s a toy for university professors to play with In the 1970s, DSP became a military advantage A very expensive technology. Only few nations were able to afford The 1980s experienced the introduction of the cost-effective DSP device Voice band modems, Hard Disk Drives and 3D Graphics were successful commercial products using DSPs The 1990s, which can best be described as the decade of consumer expansion Digital Cellular Phone can be the best example Now in 21st century DSP is every where
Impact on Society The first wave of the revolution was, and continues to be, communications Analog wire-lined world of communications became digital voice communications digital Industry moved from 2400 bps modems to 56K bps voice band modems Digital Cell Phone everywhere 802.11 DSL Bluetooth
The second wave of the revolution is entertainment Music Photo TV Games Radio Streaming Media, combination of Digital Communication and Digital Entertainment And all in the pocket same time
What could be the next wave of revolution Transportation It is a matter of time when your car will drive from point A to B without humane intervention Biomedical Science Higher quality life of a deaf man with Cochlear implant Education But there will be better uses in the classrooms as well
Why DSP is different than other Processors
Some Basic Characteristics of a DSP System Sampled data system Intensive mathematics Real time Deterministic Interrupt handling Accuracy Special hardware
Sampled Data system The signal must be sampled at twice the max frequency in the signal known as Nyquist Freqency Intensive Mathematics Most of the time in Signal Processing we do Filtering or Transforms Demand for lot of MAC operations Following is a simple filter equation For 100 tap FIR filter 100 MAC operations are needed For 25 1080p frames per Sec a 50 tap filter needs to do 2.5 million MACs per Sec for audio signal
Real Time For every input sample there is an output sample Latency, how much is the delay between input and its relative output Easiest example is audio recording. Musician has an ear piece to listen himself and others. The delay should be less than 5ms to consider it to be real time Deterministic Instruction execution time must be same every time independent of data or state of execution unit Predictability to estimate the capability of DSP to meet the demands early in the design cycle
Interrupt handling This feature was added with the concept that interrupts are entertained until the real time behavior is not effected to stick to the principle that real time systems cannot be interrupted Accurate Accuracy is signal dependent Three basic concepts to know here are SNR which is data accuracy. Simple relation is 6dB per bit Coefficient accuracy is important in filters Internal Accuracy
Special hardware To meet the needs of real time systems special hardwares that DSPs have include are Loop Hardware as most of the DSP algorithms are repetitive such FIR or FFT Data memory management as DSP algorithms often rely on present and past values of input data Saturation logic to keep the numbers inside the bound. If result exceeds the bounds maximum +ve or ve number is assigned E.g. scale all the data and coefficients to keep the results with in +1 and -1. But multiple MACs may yield the result out of this bound so the concept of guard bits in accumulator was developed Multi-MAC units executing in parallel has increased the perforamance of DSP rapidly over the last couple of decades
Bit reversal concept was adopted to increase the performance of the FFT
The evolving architecture of a DSP
When DSPs were first architected there were two generally accepted microprocessor/microcomputer architectural styles Harvard architecture (two buses architecture) Data and program memory were separate von Neumann (single bus architecture) both program and data were stored in the same memory space Neither of two fits for the DSP s demands To do Multiply in one cycle was not possible Harvard architecture was at least more suitable Little modifications were done and instructions like MPY were developed to do multiplications in one cycle Late Accumulation was also added and MAC instruction was developed
What is next in the evolution of the DSP
As first transition DSP moved from digital signal processing theory to digital signal processors Second transition is occurring now from DSP device to DSP as an enabler (DSP is no more a product in companies portfolio but is used to make products) Now there are millions of transistors in a single chip which can accommodate a SoC hosting DSP, accelerators, GPP etc Cost to produce the IC has increased many folds We are moving towards a programming world to program such systems Building such systems some may argue that its like building a complete new city from scratch like New York and where every traffic light is accurately working
When the industry introduced their first DSPs, we had about 625 instructions per sample period in a telecommunications system Performance has gone up significantly in last 20 years or so Virtually we are at a point where infinite instructions per sample are available to us
Summery: In this chapter we discussed several aspects and influences of DSP Once the theory of DSP was combined with the advancing IC technology, the DSP emerged We next discussed the revolution that the DSP began, is growing and is continuing to grow We discussed the architectural uniqueness of DSPs We discussed future and the direction DSP architectures will take Finally we discussed the drivers for those future DSPs