Announced today, Cadence's new Tensilica HiFi iQ DSP doubles compute, octuples AI performance, and cuts power by 25% over its predecessor, targeting next-gen AI-voice and immersive automotive audio.

Today, Cadence announced the production-ready IP for its sixth-generation HiFi digital signal processor (DSP). The Tensilica HiFi iQ DSP is built around a new purpose-built architecture for AI-voice and immersive audio applications. The new chip outperforms the prior generation, Tensilica HiFi 5s DSP, by 2X in compute performance, 8X in AI performance, and 40% in many audio codecs. It achieves these improvements while dropping energy consumption by 25% with most workloads. 

New AI DSP IP design for immersive and AI audio applications

Immersive and AI-powered automotive audio is one of the primary use cases for the new HiFi iQ DSP. Audio in vehicles and other consumer devices is no longer a simple one-way sound blast-out. Audio now requires generative and interpretive AI for two-way voice communications, environmental interpretation for warnings, notifications, and comfort, and high control over entertainment sound for quality and information integration. Behind those advanced capabilities are the newest generation DSP and audio-AI processors, such as the new 6th generation Tensilica HiFi iQ DSP.

Performance Improvements

All About Circuits recently spoke with Amol Borkar, group director of product management and marketing for Tensilica DSPs at Cadence about the new product.

“The design is purpose-built for next-generation audio formats, voice AI,” said Borkar, “It's built on the latest Xtensa LX8 platform and the wider execution unit that we have gives us 2X raw performance on most workloads.”

As an IP solution, Borkar relayed that the new Tensilica AI DSP comes in the form of soft IP that is production ready. It will likely be implemented within a customer ASIC or SoC, or rarely into an FPGA design, and speeds next generation customer chip time to market.

He went on to compare performance of the prior 5s DSP chip to the new 6th generation HiFi iQ DSP. 

“Out of the box the vector processing has gone up about 2X. You're going from a 128-bit to a 256-bit machine and because of that, most of the things have doubled up. In some cases, they've even quadrupled over here.” 

The charts in image below outline the performance boots and energy and area improvements over the prior chip.

Benchmark performance improvements over the Cadence generation 5 DSP predecessor. (Click on image to enlarge)

AI Features

The new DSP is able to run audio small language models (SLMs) and large language models (LLMs) on the device, without need to call back to the cloud. This means performance and capability is not compromised during network interruptions. 

It includes support for ahead-of-time (AOT)-based AI along with the SLM and LLM AI. The AI programmability aspect ensures that the device can continue to improve as AI models improve. Cadence has a compatible neural processing unit (NPU), the Cadence Neo IP, which can be paired with the DSP, or customer supplied NPU IP can be added into the design.

Key DSP Features

• 2 load units, with 2x HiFi 5s BW load units
• 5 Very long instruction word slots
• 256-bit single instruction, multiple data (SIMD) operations
• Fixed point multiply accumulate (MAC) 16 32x32, 32 32x16, 64 12x16
• AI MACs: 64/128 (16x8, 16x4,) and 128/256 (8x8, 8x4)
• 80-bit MAC accumulator width
• Dedicated AI instructions
• Optional vector floating point: FP8 16/32, BF16 16/32, FP16 16/21, FP32 8/16, FP64 4/8
• Complex floating point with ADDSUB
• Scalar Ops, CoreMark: 2, 5.51

More infromation can be found in the HiFi DSP Family product brief.

Tensilica HiFi iQ DSP. (Click on image to enlarge)

The Tensilica DSP in Use

According to Cadence, the DSP offers key automotive infotainment-oriented features. Signal processing algorithms for keyword spotting (KWS), active noise cancellation (ANC), beamforming, and automatic speech recognition (ASR) can be performed seamlessly, enabling state-of-the-art NLP functions. Additionally, for multi-stream and multichannel audio playback it facilitates rendering of 3D spatial zones and sound bubbles, delivering highly realistic listening experiences.

Delivering music is increasingly complex due to the continued introduction of new formats and new audio sources. The Tensilica HiFi iQ DSP is optimized for popular high-performance audio codecs such as: Dolby MS12, Eclipsa Audio, Opus HD, Audio Vivid, and others. Broad codec support enables broad application compatibility and reduces the design load for the end developer.

Software Application Support and Regulatory Compliance

The Tensilica IP is compatible with Cadence NeuroWeave Software Development Kit (SDK), TensorFlow Lite for Micro (TFLM), LiteRT and ExecuTorch environments for AI model execution. Ecosystem support includes multiple partner and OEM libraries, compilers, codecs, and development frameworks. The IP is designed for safety-critical applications requiring ISO 26262 functional safety (FUSA) certification. General availability is targeted for Q2, 2026 with pre-release evaluations starting in Q1.

All images used courtesy of Cadence.