New HiFi and Vision DSPs target pervasive intelligence and edge AI inference
Cadence Design Systems has expanded its Tensilica HiFi and Vision DSP families with AI enhancements and Tensilica Xtensa LX8 capabilities, designed for energy efficiency. The company has introduced four new DSPs based on the recently announced Tensilica Xtensa LX8 processor platform.
The new DSPs address increasing system level performance and AI requirements for an expanded range of applications with enhanced performance across multiple algorithms with even greater energy efficiency, claimed Cadence Design Systems. The Cadence Tensilica HiFi 1s DSP and HiFi 5s DSP are designed for audio/voice, lightweight imaging and AI applications and the Tensilica Vision 110 DSP and Vision 130 DSP are for image sensor, camera, radar and lidar applications.
“On-device and edge AI systems feature numerous diverse sensors, so any AI inference solution must be adaptable to the end application requirements,” said David Glasco, vice president of research and development for Tensilica IP at Cadence. “Designing with a system level perspective is crucial, and these SoCs must be energy-efficient, flexible and future proof to accommodate new neural networks,” he added.
The new HiFi and Vision DSPs feature AMBA AXI support for lower latency and L2 cache for increased system performance. Branch prediction enables a five to 20 per cent reduction in cycles for both the HiFi and Vision DSPs.
The DSPs are supported by Cadence’s new NeuroWeave software development kit (SDK) and can be paired with the Cadence Neo neural processing units (NPUs) to offload AI inference workloads.
The Tensilica HiFi DSPs add more capabilities for lightweight imaging and AI and can be used as additional compute resources beyond audio, either as a standalone solution or complementing the Vision DSPs and Neo NPUs.
The DSPs are designed with double-precision floating-point acceleration to provide up to 30 times better performance of popular codecs, enabling easier handling for computation requiring expanded dynamic range and precision, said Cadence.
Hardware/software co-design enhances auto-vectorisation to reduce the need for hand-optimised code and providing source code compatibility across HiFi DSPs. The company has also introduced new 8-bit operations in the HiFi 1s DSP to boost imaging and AI performance by up to two times at the kernel level, compared to the HiFi 1 DSP. In addition, the L2 cache increases system performance by up to 50 per cent.
The 128-bit Tensilica Vision 110 DSP and 512-bit Vision 130 DSP are designed to address growing sensor and AI workloads. They offload from the CPU or GPU and deliver AI performance enhancements that enable up to three times improvement for certain AI workloads and up to five times improvement on specific kernels.
In addition, Cadence said they deliver up to 20 per cent higher frequency and up to two times improvement for 16-bit, 32-bit and 64-bit floating-point performance. The company also claimed that system performance is improved through iDMA enhancements previously available only on the high performance Vision DSPs.
The DSPs also offer up to three times improvements in fast Fourier transform (FFT) for radar applications and has reduced code memory footprint due to code size improvement. The common SIMD and VLIW architecture and instruction set with Vision DSP predecessors ensure easy software migration, said Cadence.
Tensilica HiFi and Vision DSPs support Cadence’s Intelligent System Design strategy, enabling SoC design excellence. The HiFi 3z DSP, HiFi 4 DSP, ConnX 110 DSP, ConnX 120 DSP, MathX 110 DSP and MathX 130 DSP have also been upgraded to the Xtensa LX8 platform.
The Vision 110 DSP and Vision 130 DSP, and the upgraded ConnX 110 DSP, ConnX 120 DSP, MathX 110 DSP and MathX 130 DSP, are available now. The new HiFi DSPs are expected to be in general availability in December 2023.
