SensePro, a hub that processes and combines sensor data
Scalar and for floating point and integer data.
Integrating deep learning training and inference as well as parallel processing


CEVA today unveiled SensPro™, a sensor hub DSP architecture designed to handle a wide range of sensor processing and sensor fusion workloads for context-aware devices. SensePro acts as a specialized processor that can effectively handle the surge in various types of sensors.

Sensors such as cameras, radar, LiDAR, time-of-flight (ToF), microphones and inertial measurement units (IMUs) generate a variety of data types and bit rates derived from imaging, sound, RF and motion that are used to create full 3D situational awareness devices.

The SensePro architecture provides a combination of single precision and half precision floating-point operations required for high dynamic range (HDR) signal processing, point cloud generation, and voice, imaging, and deep neural network (DNN) training.

It also provides 8- and 16-bit parallel processing capacity required for voice, imaging, DNN inference processing, and simultaneous localization and mapping (SLAM).

It also provides a seamless migration path from single-sensor system designs to multi-sensor, contextually aware designs by integrating the CEVA-BX Scalar DSP.
_photo/NAPUUF44P3PV9F3YD36K.png" style="width: 600px; height: 331px;" />
SensePro can be tailored to the application using an 8-way VLIW architecture with a variety of configurations, and features an advanced deep pipeline that combines scalar and vector processing units and enables operation speeds in the 1.6GHz range at the 7nm process node. It adopts an integrated microarchitecture.

It integrates a CEVA-BX2 scalar processor for control code execution with a 4.3 CoreMark/MHz score. It adopts a wide SIMD scalable processor architecture for parallel processing and can be configured for up to 1,024 8×8 MACs, 256 16×16 MACs, dedicated 8×2 binary neural network support, as well as 64 single-precision and 128 half-precision floating-point MACs.

This provides 3 TOPS (tera operations per second) for 8x8 network inference, 20 TOPS for binary neural network inference, and 400 GFLOPS (billion floating point operations per second) for floating point operations.

Other key features include a memory architecture that provides 400 GB/s of bandwidth, a 4-way instruction cache, a 2-way vector data cache, DMA, and queue and buffer managers to offload data transfers from the DSP.

SensePro comes with a suite of software and development tools including the LLVM C/C++ compiler, an Eclipse-based integrated development environment (IDE), OpenVX API, software libraries for OpenC, and the CEVA Deep Neural Network (CENN) graph compiler.

The CDNN graph compiler includes the CDNN-Invite API for custom AI engines, CEVA-CV imaging functions, CEVA-SLAM software development kit and vision libraries, ClearVox noise reduction, WhisPro speech recognition, MotionEngine sensor fusion and SenslinQ software framework.
: center;"> ▲ SensePro DSP configuration [Table = CEVA]