CEVA SensePro2 DSP family, compared to the first generation
DSP performance is six times better, and AI inference performance is doubled.
ASIL B/D compliant, suitable for automotive use


CEVA today launched its SensPro2™ DSP family, which performs AI and DSP processing tasks related to a wide range of sensors, including cameras, radar, LiDAR, ToF, microphones, and inertial measurement units (IMUs).

The SensePro 2 DSP family delivers over 6x the DSP processing power in computer vision, over 8x the DSP performance in radar processing, 2x the AI inference capabilities, and 20% higher power efficiency on the same process node compared to the previous generation.

The new product family consists of seven vector DSP products that can be scaled to meet power and performance needs. The entry core consists of DSP and AI that require workloads with up to 1 TOPS AI performance, while the high-end core is designed to reach 3.2 TOPS.

Each product can be configured with application-specific Instruction Set Architectures (ISAs) for radar, audio, computer vision, and SLAM, as well as parallel vector computation options for floating-point and integer data types.

The SensePro2 architecture and core support hardware random error attestation for ASIL B and system error authentication for ASIL D, making it suitable for automotive applications. Performance-wise, it delivers up to 3.2 TOPS for 8x8 network inference running at 1.6 GHz, and doubles the memory bandwidth of the first generation to efficiently process data-intensive fully-connected layers.

◇ Seven CEVA SensePro2 DSP products suitable for AI/DSP tasks

The SP100 and SP50 DSPs include 128 and 64 INT8 MACs, respectively, and are smaller in die than the CEVA-BX2 scalar DSP. They deliver a 10x performance boost for DeepSpeech2 speech recognition neural networks, making them ideal for audio AI tasks such as conversational assistants, sound analysis, and natural language processing (NLP).

The SP1000, SP500, and SP250 DSPs include 1024, 512, and 256 INT8 MACs respectively, enabling optimal configurations for computer vision, SLAM, radar, and AI tasks.

The SPF4 and SPF2 floating-point DSPs contain 64 and 32 single-precision floating-point MACs, respectively, and are optimized for electric vehicle powertrain control and battery management systems. It is complemented by full linear algebra support, MATLAB vector library, and Glow graph compiler support.

SensePro2 is supported by a suite of software infrastructure to facilitate system design, including the LLVM C/C++ compiler, an Eclipse-based integrated development environment (IDE), OpenVX APIs, software libraries for OpenCL, and the CEVA Deep Neural Network (CDNN) graph compiler.

The CEVA Deep Neural Network (CDNN) graph compiler features the CDNN-Invite API with a custom AI engine, CEVA-CV imaging capabilities, CEVA-SLAM software development kit and vision libraries, radar SDK, ClearVox noise reduction, WhisPro speech recognition, MotionEngine sensor fusion, TensorFlow Lite Micro support and the SenslinQ software framework.