Linley Group: EdgeCortix IP Lowers AI Latency
Startup’s High-End DNA IP Targets Level 3 ADAS
Renowned semiconductor industry analysts The Linley Group Microprocessor Report titled "EdgeCortix IP Lowers AI Latency", concludes that EdgeCortix's unique Dynamic Neural Accelerator (DNA) architecture delivers best-in-class latency and that the...
"IP [deep-learning accelerator] achieves top scores among Level 3 ADAS (*) accelerators for batch=1 performance on ResNet-50 and Yolo v3, with equally impressive power efficiency."
The Linley Group's report provides a detailed overview of the novel EdgeCortix AI processor architecture and the forthcoming EdgeCortix DNA-A800 AI co-processor chip, showcasing more than 5x lead in AI inference latency and power-efficiency over the competition.
In real-time workloads under low-power, which are highly sensitive to latency (response time) and require streaming data or small batches, EdgeCortix DNA-A800 with 54 TOPS of computing in single-core configuration, delivers breakthrough batch-size 1 throughput of 2500 IPS (images per second) for ResNet-50 v1.5, and 555 IPS for Yolo v3, while consuming under 10 Watts of power. This translates into more than 15x IPS/Watt (power-efficiency) advantage over competing architectures.
ResNet-50 v1.5 is an inference benchmark for image classification and is part of a suite of MLPerf standards for measuring the performance of deep learning hardware accelerators.
Yolo v3 is the state-of-the-art inference benchmark for real-time object detection in images and videos.
* ADAS is the abbreviation for Advanced Driver-Assistance Systems in automobiles