Speeds deep learning inference by 5x compared to TensorFlow on NVIDIA GPUs

Bangalore, 27 March 2018 – MathWorks today announced that MATLAB now offers NVIDIA TensorRT integration through GPU Coder. This helps engineers and scientists develop new AI and deep learning models in MATLAB with the performance and efficiency needed to meet the growing demands of data centers, embedded, and automotive applications.

MATLAB provides a complete workflow to rapidly train, validate, and deploy deep learning models. Engineers can use GPU resources without additional programming so they can focus on their applications rather than performance tuning. The new integration of NVIDIA TensorRT with GPU Coder enables deep learning models developed in MATLAB to run on NVIDIA GPUs with high-throughput and low-latency. Internal benchmarks show that MATLAB-generated CUDA code combined with TensorRT can deploy Alexnet with 5x better performance than TensorFlow and can deploy VGG-16 with 1.25x better performance than TensorFlow for deep learning inference.*

“Rapidly evolving image, speech, sensor, and IoT technologies are driving teams to explore AI solutions with better performance and efficiency. In addition, deep learning models are becoming more complex. All of this puts immense pressure on engineers,” said David Rich, director, MathWorks. “Now, teams training deep learning models using MATLAB and NVIDIA GPUs can deploy real-time inference in any environment from the cloud to the data center to embedded edge devices.”

To learn more about MATLAB for deep learning, visit: mathworks.com/solutions/deep-learning.html

* All benchmarks were run on MATLAB R2018a with GPU Coder, TensorRT 3.0.1, TensorFlow 1.6.0, CUDA 9.0 and cuDNN 7 on an NVIDIA TITAN Xp GPU in a Linux 12 core Intel ® Xeon® E5-1650 v3 PC with 64GB RAM