Processor vendors and system designers share a common concern: how to make sure their products meet customer needs. For processor vendors, a key challenge is to design architectures with enough performance to meet the demands of current and anticipated applications while staying within acceptable power and cost constraints, and enabling good software developer productivity. Fundamentally, processor designers need to bring together the demands of algorithm workloads together with the capabilities of processor architectures.
A recent BDTI client engagement illustrates how a deep understanding of algorithms and architectures can result in the right design. A leading worldwide embedded processor vendor needed to demonstrate to its customers that its processor roadmap included devices capable of powering increasingly demanding deep learning-based applications to detect objects from visual data. As an example of a challenging workload, BDTI engineers selected the Faster R-CNN neural network – a state-of-the-art object detection network that combines a region proposal network (RPN) with a convolutional neural network (CNN).
BDTI carefully analyzed the processing requirements of the network, and then modified it to create a streamlined version suitable for an embedded processor. BDTI determined how each block of the algorithm could take best advantage of the processor’s computational resources, which include CPU, GPU, DSP, and specialized processor cores. BDTI then mapped the algorithm onto the processor architecture, optimizing for best use of both the processing units and the memory architecture of the SoC. BDTI delivered a clear analysis of the expected performance of the algorithm, along with a modeling tool that enables the client to show how algorithm performance changes as the number and type of processing units are changed, and as parameters of the algorithm (such as image resolution) are changed. With this information, the processor vendor was able to clearly articulate how the SoCs in its roadmap will meet the anticipated needs of its customers.
The increasingly demanding algorithms employed in a growing number of smart products for consumer, enterprise, industrial, and other markets place unique demands on processors. Processor designers benefit by carefully considering the algorithms that their customers will run on their chips. If you are designing a processor for a demanding computer vision or deep learning application, or if you are a system designer looking for a processor to power your design, contact Jeremy Giddings (giddings@BDTI.com) for a confidential discussion of how BDTI can help.