Popular since the 1970s for general-purpose computing and supercomputing, floating point is increasingly in demand for compute-intensive digital signal processing applications like sensor fusion, equalization, vision processing, radar processing, and multi-antenna MIMO. For the type of signal processing required by these applications (think Kalman filtering, IIR filters, and multiple signal classification, or MUSIC, algorithms in radar), floating point provides the best performance and sometimes lower power consumption. Today, Cadence launched the latest optimizable Tensilica processor, the Xtensa LX7 (Figure 1), which increases floating-point scalability from 2 to 64 FLOPS/cycle for a broad portfolio of DSPs that are used in many markets, including: convolutional neural networks (CNNs), autonomous cars, and internet of things (IoT) devices. What else is driving the increased use of floating point? Here are five key trends, with a look at how the Xtensa LX7 is addressing each. Lower latency : Compute-intensive applications typically need to process large volumes of data in realtime, so the shorter the delay between the input and output, the better for smooth, responsive operation. With the Xtensa LX7 release, an integrated direct memory access (iDMA) controller option offloads memory-to-memory data transfer operations from the processor, delivering lower latency along with less system bus bandwidth usage and lower power. Dynamic range and greater precision : With their ability to represent very small numbers as well as very large numbers, floating point delivers the precision that compute-intensive applications need. The Xtensa LX7 features click-box IEEE 754-compliant single-, double-, and half-precision floating-point options. Ease of programming and software resources : Floating-point DSPs have long been recognized for being easy to program. The Tensilica Xtensa Software Developer’s Toolkit provides an extensive set of code generation and analysis tools that help accelerate the application software development process. There are simulation models, RTOS ports, optimized C libraries, third-party tools compatibility, and more. Lower power : Many of today’s compute-intensive applications are also power sensitive. IoT devices such as battery-operated fitness trackers and smart home security systems consist of components with small form factors and must run efficiently. All Tensilica processors are based on an efficient, low-power 32-bit controller architecture and support low leakage power and dynamic power savings such as semantic and data power gating. A power shut-off feature lets you power off the processor completely. Time to market : Fast delivery to market is critical for most any design. With its ease of programming and software tools, you can quickly optimize a Tensilica processor for your specific application. Since the processor is fully supported by automatic hardware and software generation, you can easily future-proof and add performance to your designs through software programming and differentiate your designs through unique processor implementations targeted to your application. As an application example, consider a typical radar signal processing module. Such modules rely on direction of arrival algorithms like MUSIC, which require computing eigenvalues and eigenvectors of matrices, roots of polynomials, and kernels sensitive to data precision and hard-to-predict dynamic range. For these characteristics, floating point is ideal. The Xtensa LX7 is the basis for the new Tensilica Vision P6 DSP for image, computer vision, and CNN processing, the new Tensilica Fusion G3 DSP for multi-purpose fixed- and floating-point applications, and the ConnX BBE DSPs for baseband and radar applications. Read the Xtensa LX7 datasheet for more details. Christine Young
↧