I'm sure you've noticed that there is a lot of talk about 5G in the air. Well, "in the air" is the one place it isn't, since it is a new standard that is being brought to market over the next few years. There are a lot of basestations to be created and installed before handsets will be truly useful.

There's also a lot of talk about autonomous driving. In fact, at CES in January, there was a huge hall dedicated to "transportation" and the focus of transportation is electronics (at least at CES—the gear-head shows might be different).

These new capabilities are going to require new processors, and today at a press conference at embedded world in Nuremberg, Cadence announced the latest member of the Tensilica family, targeted at these nascent markets, 5G and automotive. This is the Tensilica ConnX B20 DSP. Notwithstanding that the markets are not yet fully developed, many SoCs are being designed this year.

That may sound odd, since automotive and mobile are pretty separate markets, notwithstanding that cars are obviously mobile. But one thing the seeming disparate trends in these markets have in common is the increased need for high-performance signal processing. Let me explain.

5G Communications

First, 5G. Every generation of mobile standard has made the same trade-off: take the big increase in available computational resources delivered by a mixture of Moore's Law and more sophisticated processor architectures, and use it to get more efficient use out of the air interface. That's mobile-speak for getting more data transmitted using the same amount of radio bandwidth, and with less energy. It is analogous to listening to someone talking in a noisy environment like a party—it takes more of your brainpower to understand them if they talk quietly, but they are "wasting" less energy transmitting. The available spectrum is a truly limited resource for mobile, although we have to be miserly with battery power to take advantage of it, especially at the handset end.

With 5G we will have extensive use of multiple antennas (known as MIMO) and beam-forming. Since a handset knows where the basestation is, it can optimize power by directing the signals at the receiving antennas, and as a result not spraying wasted radio signals in all the wrong directions. This is known as beamforming. The bandwidth will also be increased, especially in the highest frequency 5G-only band, known as mmWave, which will provide very high bandwidth and low latency over short distances (oxygen absorbs radio waves at those frequencies, so distances are limited to about 1000 feet).

Autonomous Driving

There are a number of aspects to autonomous driving, but a key one is to make the car aware of its surroundings. In the longer term, it will be able to drive itself autonomously. In the shorter term, under the umbrella term ADAS (for Advanced Driver Assistance Systems) it will do simpler tasks like lane following and automatic emergency braking (AEB will be mandatory equipment in Europe by 2021 and the US by 2022). Obviously, both lane following and AEB require the car to be able to "see". Here's a video from just last week of someone's life maybe being saved by AEB (10 seconds):

Today, on cars that are in commercial production, there are two ways of making the car aware of its surroundings: radar and cameras.

You probably already know this, but just in case not, radar works by sending out a radio-frequency "chirp" which bounces off an object and then the returning signal is processed to work out what is out there, most obviously by timing the speed-of-light delay to get the distance (called RFFT), but also Doppler for speed, called DFFT, and angle-of-arrival or AoA, for position. Today, radar is great for seeing in bad weather (and in the dark, it can see further than headlights) but has insufficient resolution to distinguish distant objects. The trend is to increase the amount of processing that is done on the radar signals to improve the resolution, using more antennas.

On the research "self-driving-cars" that are not yet in commercial production, you see lidar (which is basically radar using lasers, sending out pulses of light). That's the big spinning thing on the roof. Those units cost about as much as the vehicle underneath them, so the cost has to come down for lidar to be commercially feasible. The focus of doing that is to use solid-state lasers, along with approaches like phased arrays, or MEMS approaches such as micromirrors. Then you can lose the big spinning thing. But that again puts heavier compute load on the processors to derive good information from lower resolution images.

Common Requirements

Cadence has an existing line of processor targeted at the infrastructure and automotive markets (there'll be a test at the end of this post to see if you can remember these catchy names) the ConnX BBE64EP, BBE32EP, and BBE16EP. These are high-performance, with full hardware support for complex numbers, and other features needed for these types of complex DSP algorithms. But for the next generation of cellular, radar, and lidar, this is not enough compute horsepower.

As I said above, it sounds a little counter-intuitive, but 5G communications, high-resolution automotive radar, and miniaturized automotive lidar all have a (roughly) common set of requirements: faster processing, higher accuracy, higher energy efficiency. Getting more specific, they require:

• Complex numbers: Heavy duty DSP algorithms are implemented with complex numbers.
• 32-bit floating point: 16-bit lacks the accuracy required.
• Higher clock speed: Obviously higher performance, but the power needs to be kept under control.
• More parallelism: Higher performance.
• ISA optimized for the application: Focus the limited power on doing a lot of real work per instruction, and keep the overhead of programmability down.

ConnX B20 DSP

Today, at embedded world in Nuremberg, Cadence announced the next-generation Tensilica ConnX B20. This is built on the deeper Tensilica pipeline architecture. For some detail about that, see my post, A New Era Needs a New Architecture: The Tensilica Vision Q6 DSP. The clock frequency on a 16nm process is 1.4GHz or greater. For communication applications, the performance can be as much as 30X higher than the ConnX BBE32EP. For radar and lidar applications, up to 10X faster. Obviously, the B20 is the highest performing ConnX DSP.

A block diagram of the B20 is above. The key points of the new processor are:

• 32-bit FP vector MAC option
• Deeper pipeline and higher clock speed
• New HPX and SPX options (extended floating point)
• More parallelism:
  • 32-bit vector MAC option adds 2X 16x16 operations
  • SPX option doubles SP performance
• Optimized ISA:
  • Communications option for forward error correction (FEC) acceleration (needed for 5G)
  • Improved 16-bit sin/cos/exp, FFT scaling
• Software compatible with other members of the family

It is certainly possible to build a processor with lots of compute resources that does well on benchmarks but performs poorly on real-world application code. The B20 has specifically been targetted at 5G, and automotive radar/lidar. It performs very well with:

• Up to 30X improved performance in the appropriate parts of 5G communications applications
• Up to 10X improved performance in automotive applications

More Details

For full details, see the ConnX DSP product page.

Sign up for Sunday Brunch, the weekly Breakfast Bytes email.