Always Listening, HiFi 1 DSP

28 Oct 2021 • 4 minute read

woman using earbuds Today, at the Linley Processor Conference, we announced the latest addition to the HiFi DSP family, the Tensilica HiFi 1 DSP. This joins the existing products: HiFi 3 DSP, HiFi 3z DSP, HiFi 4 DSP, and HiFi 5 DSP. As you might guess from the number, the HiFi 1 DSP slots in at the low end of the spectrum. It is smaller and lower power than the other HiFi processors...it is designed specifically for always-on applications. Like the other family members, it is equipped for AI so that it can do things like listen for a wake-word ("Alexa", "OK Google", "Hey Siri").

See my post HiFi DSPs - Not Just for Music Anymore for details of the HiFi DSP portfolio and my post "Alexa, What Is HiFi 5?" for details on that DSP.

However, there are other applications than always-on where there are strict power or size limitations. An obvious one is wireless earbuds. These are not strictly always-on, but some people listen to music for many hours continuously. The chips cannot be physically large, nor can the battery. But battery life, how long you can listen to music without recharging, is a key feature of the device. So encoding and playback has to be processed by an extremely efficient DSP.

smartphone in pocket with screen powered on Another application is detecting the state of the phone (at ear, face-up, face-down in pocket, etc) and managing whether the screen and its back illumination are powered on or not. This requires an always-on sensor to run at very low energy levels. It can waste a lot of energy having the screen powered on when it is not being interacted with, as in the photo.

A similar always-on sensing issue is for laptop data security. Shutting a laptop down and manually locking it is not always feasible, but an unattended system still needs to be secured, and preferably automatically.

Artificial Intelligence (AI) is increasingly used for context awareness algorithms for functions like noise suppression, or sound analytics, and so require not just a low-power DSP but a low-power DSP with neural network inference capability.

requirements for always on

The diagram above shows the applications for always-on (or mostly-on) use cases, and these feed into PPA requirements for the processor (and the whole SoC).

Tensilica HiFi 1 DSP

The HiFi 1 is the smallest Tensiilica HiFi DSP. It is energy and cycle-count optimized for Bluetooth (BT) and Bluetooth Low Energy (BLE) voice and music codecs such as LC3/ There are special neural-network instructins for keyword spotting and light neural network algorithms. There is an optional vector floating-point unit (VFPU) for rapid "MATLAB to optimized DSP" algorithm porting. It is software-compatible with the other HiFi DSPs and, like them, can also be customized using the TIE language (Tensilica Instruction Extension...see my post Custom Instructions in Tensilica: Wearing a TIE Makes You Smarter for more details).

This table shows the detailed features of the HiFi 1 DSP:

hifi 1 specifications

There are innovations for efficient VLIW slot utilization, with the VFPU (which is optional) able to operate in both slots, and a high Coremark score for efficient control (integer) code. There is a lightweight special ISA for neural-network operations, including efficient load/store of 8-bit data and efficient dot-product and convolution. Depending on the technology node obviously, but in advanced nodes, it can be clocked to over a gigahertz.

Above is the block diagram of the HiFi 1 DSP.

How good is the performance of the HiFi 1 DSP? Let's compare it to the bigger HiFi 3 DSP, the most popular DSP for Bluetooth codecs. It is 16% smaller (11% with the optional VFPU), uses 18% fewer cycles, and 14% less energy. On an "OK Google" keyword detection, it uses 60% fewer cycles and 62% less energy.

Several applications of HiFi 1 DSP use it to decide when the main system needs to be woken up. For example, in data security in laptops, the HiFi 1 DSP can use efficient time-of-flight (TOF) to detect a person, and then the main system can be woken to identify the owner with the camera (which would be too expensive from an energy point of view to do all the time). This can reduce access time by turning a laptop back on again when the owner approaches.

For audio applications, such as earbuds, the HiFi 1 DSP can be used as the main DSP providing ultra-low energy audio playback or voice calls (remember them?). Alternatively, it can be used as the always-on DSP that then wakes up the main system when user intent is detected (via a wake word or pressed button). For AI use-cases, the HiFi 1 DSP can be paired with the new NNE110 neural-network processor giving AI boost but still at low energy levels.

hifi 1 with NNE110

The HiFi DSP Portfolio

Let's wrap up with a look at the whole HiFi DSP portfolio, from HiFi 1 DSP at the low end up to HiFi 5 DSP for much higher performance (at a cost in power and area, of course):

hifi portfolio This portfolio has been very successful, with over 130 licensees, 1.5B cores shipping annually, and over 300 audio software packages provided by over 160 ecosystem partners. It is #1 in the marketplace as the audio DSP of choice.