Cadence’s Tensilica DSP is a highly configurable embedded software platform for implementing AI at the edge with highest energy efficiency and best-in-class inferences per second for resource-constrained designs. The Tensilica Vision DSPs in particular can be used with camera-based vision, radar, and lidar systems for streamlined AI imaging and inferencing on a host CPU/GPU.
Tensilica Vision DSP can also be used to create a Large Vision Model (LVM) AI that can run on-chip. This enables direct text-based searches for classified objects within an image, for example.
Several varieties of DSP are available depending on the processing requirements of the use case. Key application areas include AR/VR headsets, automotive, mobile devices, drones and robots, assembly lines, and surveillance cameras.
Key features of the Tensilica DSPs include:
- 5-way VLIW (Very Long Instruction Word) SIMD (Single Instruction, Multiple Data) (128 to 1024-bit) machine vision
- Data-level parallelism with SIMD MACs: 4×4, 8×16, 16×16, 32×32
- Up to 1024 (8×8) MACs
- 8/16/32-bit fixed-point processing and complex fixed-point support
- Half-, single-, and double-precision floating-point units (FP16/32/64)
- Optional performance boost for FFT (Fast Fourier Transform) and SLAM (Simultaneous Localisation And Mapping)
- Support for user-defined instructions
- Integrated DMA
- ISO 26262 hardware and software safety compliance
Cadence provides extensive framework support for getting started using its AI vision DSPs, including highly optimised imaging libraries and the ability to migrate existing codebases using OpenVX, OpenCL, and Halide. Tensilica also supports the ONNX and GLOW AI networks for the most cutting-edge innovations in vision AI.
Fill out the form below to get connected with Cadence for more information and the first steps to evaluating this technology for use in your next commercial design.