Fusion core targets voice-activated devices
Cadence Design Systems has launched a processor core aimed at ‘always on’ signal-processing applications such as voice detection and recognition for wearables.
The company has derived the Tensilica Fusion core from its existing HiFi 3 DSP, lowering the resolution of the parallel multiplier engine to improve energy efficiency on voice-detection algorithms and adding optional bit-manipulation instructions to support common low-power wireless radio protocols such as Bluetooth Smart.
Gerard Andrews, senior product manager for audio and voice DSP products in Cadence’s Tensilica IP group, said: “We believe voice could be more important as a user interface across a broad spectrum of these devices.”
The Fusion can perform both DSP and microcontroller functions, so could act as a standalone processor or sit alongside another applications processor used for higher-level functions. Andrews said the core has RTOS support.
Quad MAC for wireless
Like the HiFi 3, the Fusion has a multiply accumulate (MAC) unit. However, it is narrower than the HiFi 3, supporting single 32 x 32bit or 24bit MAC operations or dual 32 x 16bit or 16 x 16bit MACs. To fit the typical requirements of voice algorithms such as Sensory’s Truly HandsFree and narrowband wireless protocols, an optional quad-16bit MAC configuration is possible. An AES engine can also be added.
The Fusion supports single-precision floating-point operations. “A lot of the sensor-fusion companies have great libraries but they have often not been converted to fixed-point arithmetic,” Andrews said.
A further energy optimization is the loop buffer, a memory block of configurable size that stores the core DSP kernel code for a voice detector or similar algorithm. While the processor is running the code in that small kernel the full program memory block can be powered down or kept in a low-power drowsy mode until it is needed.