Cadence Tensilica HiFi iQ DSP IP is the sixth generation of the company’s HiFi DSP family found in countless SoCs with audio features, and is based on a new architecture designed for next-generation voice AI and immersive audio applications.
Compared to the Tensilica HiFi 5s DSP, the HiFi iQ DSP offers twice the compute performance, 8x higher AI performance, and more than 25% energy savings for most workloads, while delivering over 40% performance uplift on several audio codecs. The company claims it can act as an all-in-one AI processor for voice AI applications running popular SLMs and LLMs on the DSP itself.
Block diagram
Tensilica HiFi iQ DSP key specifications:
- Load Units – 2
- VLIW Slots – 5
- Vector operations (SIMD) – 256-bit
- Accumulator Width – 80-bit
- Fixed Point MACs
- MACs for AI – Up to 128 16×8, up to 256 16×4
- Optional instructions for AI (improved from HiFi 5s)
- Optional Vector Floating Point – FP8 (conversion), BF16 (conversion), FP16, FP32, FP64
- Complex Floating Point with ADDSUB
- Auto-vectorization improved from HiFi 5s for easy programming; it removes the need for manual optimization to speed up firmware development and time to market
Cadence states that the Tensilica HiFi iQ DSP can support advanced audio codecs such as Dolby MS12, Eclipsa Audio, Opus HD, and Audio Vivid with higher efficiency thanks to architectural enhancements and increased computational capability.
The new DSP also enables signal processing algorithms for keyword spotting (KWS), active noise cancellation (ANC), beamforming, and automatic speech recognition (ASR) for Natural Language Processor (NLP) functions. Furthermore, multi-stream and multi-channel audio playback facilitates rendering of 3D spatial zones and sound bubbles. The HiFi iQ DSP will also be suitable for SoCs designed for safety-critical applications compliant with ISO 26262 functional safety (FUSA) certification. Support for cache-coherent multicore configurations is also in the works.
The new Tensilica DSP is compatible with the Cadence NeuroWeave Software Development Kit (SDK), TensorFlow Lite for Micro (TFLM), LiteRT, and ExecuTorch environments for AI model execution. The company also highlights software libraries, compilers, codec packages, and frameworks from partners and OEMs.
Design flow for SoCs with Tensilica DSP