Get email delivery of the Cadence blog featured here
Only last week, in Are General-Purpose Microprocessors Over? I wrote about how general-purpose processors are never going to get any faster and so the system architecture of the future will consist of a general-purpose processor (probably multicore) along with a number of special-purpose processors. A lot of those special-purpose processors will be digital signal processors (DSPs) of one sort or another.
Another driver of the importance of DSPs is that the world is analog. Historically, this has meant complex analog interfaces. However, in a modern process, analog designs don't shrink much, if at all, or may simply be impossible to design. But digital gates and DSPs halve in area each process node, and so the balance of power shifts in favor of handling signals with only the minimum amount of analog and doing everything else in the digital domain with DSPs.
DSPs hit the sweet spot between simply programming the general-purpose processor, which won't be fast enough and also too power-hungry, and designing at the RTL level, which doesn't have the flexibility required, especially in most application areas for DSPs where new standards appear and old standards are revised. The general-purpose processor is too hot, RTL is too cold, and a DSP is just the right temperature for the signal processing porridge.
For some applications, such as audio, DSPs can get even more specialized for just that application. Manufacturers of smartphones want you to be able to listen to music all day without depleting the battery, and so low power is king (audio is not especially demanding of processor performance). Tensilica HiFi DSPs are the market leader in this domain. However, many systems require more general-purpose signal processing. For this the Tensilica Fusion G series DSPs are a good solution.
It is a solution that just got better. Today, Cadence announced the newest member of the family, the Tensilica Fusion G6.
As you might guess, it is similar to the Fusion G3 DSP but higher performance. Like all Tensilica processors, it is built on top of the underlying Tensilica Xtensa LX processor, and the TIE technology that allows user-defined optimizations and special instructions. The table below shows the differences between the Fusion G6 DSP and the Fusion G3 DSP.
The short version is that almost all the important specifications have doubled from the Fusion G3 DSP to the Fusion G6 DSP. More MACs, wider caches, wider datapaths. The Fusion G3 and G6 DSPs are targeted at general-purpose applications where raw DSP power is required. It is especially attractive when the details of the application are not, or cannot be, clearly known in advance. Either the DSP is used by the system for multiple purposes at different times, or the decision about which DSP to put into the SoC has to be taken before all the requirements are fully nailed down. Multiple algorithms can be run on the same processor with the same instruction set architecture (ISA). It is very flexible, supporting both fixed- and floating-point datatypes, and efficient real time control.
The performance that results is shown in the graph above (versus the Fusion G6 DSP, which is at 100%). The Fusion G3 DSP requires around half the cycles of competitive processors, and the Fusion G6 DSP is only half as many as the Fusion G3 DSP.
There are many application areas that can make use of a general-purpose DSP like the Fusion G6 DSP, from IoT to automotive, from mobile to healthcare. Within those markets, the diagram above shows some of the most applicable functions that can be implemented using the Fusion G DSP family.
Every DSP selection problem is unique, of course, but common features that are driving adoption of the Fusion G DSPs are:
At CDNLive Munich today, Cadence's Pranava Tummala will present on MIMO Radar Signal Processing on Tensilica DSPs and will show how to do this on the Fusion G3 and G6 DSPs. MIMO stands for multiple-input-multiple-output. In the context of mobile phones, this means using multiple antennas, often in different bands, to get high throughput. In the context of radar, it means using multiple antennas to increase the resolution. There is a lot of signal processing: range FFT (RFFT), Doppler FFT (DFFT), constant false alarm rate (CFAR), angle of arrival (AoA). Together, these allow for the vehicle to detect other vehicles, eliminate noise, and determine the position and speed of the other vehicles.
The presentation goes into depth on how to map these radar processing algorithms onto various system architectures: the single-core Fusion G DSP, single-core Fusion G6 DSP, and four-core Fusion G6 DSP. The basic tradeoff in radar is that higher resolution rates require higher processing power, using a Fusion G6 DSP instead of a Fusion G3 DSP and adding additional cores. The above chart shows the effect of increasing the size of the datacube (channels x doppler x range) on the performance requirements for the four basic radar signal processing operations.
For more detaiis, if you are at CDNLive later today then attend the presentation. Otherwise, the presentation will be available online in few weeks.
Get more details about the Tensilica Fusion G3 and G6 DSPs, including a product brief.