DSPs: What They Are and Why They Are Used

By, Amir Majidimehr

Let’s take a look at the typical spec sheet for an Audio/Video processor, in this case an Anthem D2v:

“Two dual-core digital signal processing (DSP) engines, our own DSP design, offer a total of 800 MIPS to allow decoding of the new HD audio standards: Dolby Digital Plus, Dolby TrueHD, DTS-HD High-Resolution Audio and DTS-HD Master Audio. More than enough processing power to handle even the most complex program material with matchless precision.”

Some terms in there should make sense to you such as the compressed audio standards used in DVD and Blu-ray disc (DTS, Dolby, etc). But what on earth is a DSP?
Analog Devices Shark DSP

DSP stands for Digital Signal Processor. As the last word implies, this is a device that is akin to the processor in your computer. Its job is to read instructions, work on data, and generate results. An engineer with knowledge of the above tasks creates a program for it that it loads and runs as instructed. In the case of above examples, it will decode a Dolby TrueHD compressed stream into PCM audio samples ready to be output to the DAC (Digital to Audio Converter).

The above may sound like what is also going on in your computer but DSP’s function and design are optimized for a different task. To understand that we need to step back and talk about what type of processing we need to perform in audio domain.

Take the functions of the Anthem AV processor. In addition to the above decoders it also performs functions such as room equalization and bass management. How these “algorithms” (methods) work is beyond the scope of this article. What is important to understand about them is that we can reduce them into a sequence of math operations. For example there is math function where we multiple one number by another and keep accumulating the result. Such a function can then be used to implement a filter to for example route high frequencies to your speakers and low ones to the subwoofer. This operation is so common that it has its own name called “MAC” which stands for multiply-accumulates.

Now imagine having to figure out the above sum for every audio sample. If you are playing most movie soundtracks, the sampling rate is 48,000/sec meaning there are 48,000 samples per second. For every audio sample we may be performing hundreds of arithmetic operations. Since our system is “real time” we need to keep up with the incoming audio rate and compute our operations in the same allotted time. So in my hypothetic example of 100 operations per audio sample, we need to perform nearly five million operations per second (48,000 * 100). This is per channel! And the number scales up with sampling rate. If we have 7 channels of audio and at 192 KHz sampling rate, the number of operations climbs to 33 million/second. If this doesn’t impress you, remember, this is just for one function the processor is performing. In real life, it would be performing cascaded functions of many, decoding audio, performing room EQ, bass management, etc.
Shark DSP Architecture

DSPs are purpose built devices that are optimized to perform repeated mathematical operations. In the above example, the same bit of math is applied to every audio sample repeatedly millions of times. Having this knowledge allows one to build a processor that is a) faster, b) cheaper, c) lower power or all of these. Why? Well, we know certain things about how the processor will ultimately be used. For example, the values used to multiply the audio samples are a small set that are used over and over again. So what we can do is to have very high-speed memory inside the processor that keeps this data. This memory would be instantly accessible to use as opposed to traditional computer systems where data is normally kept in slower main memory.   Similarly a single instruction can tell the DSP to perform a sequence of math functions so the overhead of telling it what to do relative to the work it has to do is very low.

Traditionally DSPs had a major leg up over processors in your computers when it came to arithmetic operations. They could use all of their silicon real estate for math functions allowing them to have what we call very high “peak” computational power.   Indeed the “super computers” of the past were built the same way. You don’t hear much about those super computers any more (at least not the traditional ones). For the same reasoning, this advantage of DSPs in sheer power is mostly lost as compared to modern CPUs.

How did that happen? Well, CPU designers stole the DSP technique and ran with it! It used to be that CPUs only knew how to fetch and execute one math operation at a time. That created a loss of efficiency as instruction after instruction had to be executed to perform the same math operation (on different data). Some 20 years ago Intel started a trend of adding so called “SIMD” extensions to mainstream computer processors. SIMD stands for Single Instruction, Multiple Data. As the name implies, one instruction can tell the CPU to perform multiple ones which is precisely the type we need to match what the DSP does. While the solution is not quite as optimized as it is with DSPs, it comes awfully close.   And with it, your mainstream CPU now can perform these specialized math functions with very high peak speed.

So why do we do we not use a computer CPU in our AV processor? One answer is the cost. The CPU and its associated circuitry cost considerably more than a DSP since they have to be good at many things and not just arithmetic.

Another reason is power consumption. All else being equal, a simpler circuit uses less power. DSPs are streamlined to perform a set of mathematical functions. This translates into less digital circuits “toggling” and with it, less power consumed.   The Anthem processor is not battery operated so you might think it shouldn’t care about power but it does. With so much functionality in today’s AV processors, we need to manage the total amount of heat generated and avoid having fans and such that many of our computers sport. High power can also translate into more cost in the form of heat sinks and larger power supply.

Is there a down side to DSPs? Technically yes. They “hate” doing work that is not repetitive. Telling it to do something if a number is less than another but something else if not, sharply reduces its performance. For this reason, they make lousy general purpose computers and hence the reason your computer doesn’t use them.

By the way, there is even a more optimized way to perform these functions and that is through purpose built “hardware.” That is a circuit that is designed to do one and only one thing. It goes by the names “FPGA” or “ASIC.” Video decoding for example is done using these techniques as using DSPs or CPUs to do the same would be much more expensive. The advantage DSPs have is that they are flexible. As you see implied in the Anthem spec, we can load up the DSP with different audio decoder programs and have it perform that function. If we had done the same with FPGA/ASIC, we would have needed to have a chip with all of those functions simultaneously implemented since there is no program to load and tell it to run. Since we never need to decode Dolby and DTS at the same time this approach would have been a waste.   This is why you always see a DSP in an AV processor dedicated to audio tasks.

So another way to think of DSPs is as programmable hardware. It is good for semi-fixed function tasks that typify audio processing.

I should mention that programming DSPs is a difficult thing. It is specialized skill that requires knowledge of the task needing to be done and the best way to get it to be a long sequence of math operations as anything else would run too slow. For this reason, I personally have a lot of respect for those who can do it well!

Back to Articles