NVIDIA TensorRT
NVIDIA TensorRT™ is an SDK for high-performance deep learning inference. It includes a deep learning inference optimizer and runtime that delivers low latency and high-throughput for deep learning inference applications. TensorRT-based applications perform up to 40x faster than CPU-only platforms during inference. With TensorRT, you can optimize neural network models trained in all major frameworks, calibrate for lower precision with high accuracy, and finally deploy to hyperscale data centers, embedded, or automotive product platforms.
TensorRT is built on CUDA, NVIDIA’s parallel programming model, and enables you to optimize inference for all deep learning frameworks leveraging libraries, development tools and technologies in CUDA-X for artificial intelligence, autonomous machines, high-performance computing, and graphics.
TensorRT provides INT8 and FP16 optimizations for production deployments of deep learning inference applications such as video streaming, speech recognition, recommendation and natural language processing. Reduced precision inference significantly reduces application latency, which is a requirement for many real-time services, auto and embedded applications.
(Click to Zoom)
You can import trained models from every deep learning framework into TensorRT. After applying optimizations, TensorRT selects platform specific kernels to maximize performance on Tesla GPUs in the data center, Jetson embedded platforms, and NVIDIA DRIVE autonomous driving platforms.
To use AI models in data center production, the TensorRT Inference Server is a containerized microservice that maximizes GPU utilization and runs multiple models from different frameworks concurrently on a node. It leverages Docker and Kubernetes to integrate seamlessly into DevOps architectures.
With TensorRT developers can focus on creating novel AI-powered applications rather than performance tuning for inference deployment.
TensorRT Optimizations and Performance
Weight & Activation Precision Calibration
Maximizes throughput by quantizing models to INT8 while preserving accuracy
Layer & Tensor Fusion
Optimizes use of GPU memory and bandwidth by fusing nodes in a kernel
Kernel Auto-Tuning
Selects best data layers and algorithms based on target GPU platform
Dynamic Tensor Memory
Minimizes memory footprint and re-uses memory for tensors efficiently
Multi-Stream Execution
Scalable design to process multiple input streams in parallel
TensorRT dramatically accelerates deep learning inference performance on NVIDIA GPUs. See how it can power your inference needs across multiple networks with high throughput and ultra-low latency.
Widely Adopted
Integrated with All Major Frameworks
NVIDIA works closely with deep learning framework developers to achieve optimized performance for inference on AI platforms using TensorRT. If your training models are in the ONNX format or other popular frameworks such as TensorFlow and MATLAB, there are easy ways for you to import models into TensorRT for inference. Below are few integrations with information on how to get started.
TensorRT and TensorFlow are tightly integrated so you get the flexibility of TensorFlow with the powerful optimizations of TensorRT. Learn more in the TensorRT integrated with TensorFlow blog post.
MATLAB is integrated with TensorRT through GPU Coder so that engineers and scientists using MATLAB can automatically generate high-performant inference engines for Jetson, DRIVE and Tesla platforms. Learn more in this webinar.
TensorRT provides an ONNX parser so you can easily import ONNX models from frameworks such as Caffe 2, Chainer, Microsoft Cognitive Toolkit, MxNet and PyTorch into TensorRT. Learn more about ONNX support in TensorRT here.
TensorRT is also integrated with ONNX Runtime, providing an easy way to achieve high-performance inference for machine learning models in the ONNX format. Learn more about ONNX Runtime - TensorRT integration here.
If you are performing deep learning training in a proprietary or custom framework, use the TensorRT C++ API to import and accelerate your models. Read more in the TensorRT documentation.
“In our evaluation of TensorRT running our deep learning-based recommendation application on NVIDIA Tesla V100 GPUs, we experienced a 45x increase in inference speed and throughput compared with a CPU-based platform. We believe TensorRT could dramatically improve productivity for our enterprise customers.”
— Markus Noga, Head of Machine Learning at SAP
“By using tensor cores on the V100, the most recently optimized CUDA libraries and the TF-TRT backend we were able to speed up our already fast DL network by a factor of 4x”
— Kris Bhaskar, KLA Senior Fellow, VP AI initiatives, KLA
“Criteo uses Nvidia's TensorRT over T4 cards to optimize its deep-learning models for faster inference on GPUs. Now, removing inappropriate images over billions of them is 4 times faster. It also consumes half less energy.”
— Suju Rajan, SVP Research, Criteo
NVIDIA TensorRT Inference Server
NVIDIA TensorRT Inference Server simplifies deploying AI inference in data center production. It’s an inference microservice for data center production that maximizes GPU utilization and seamlessly integrates into DevOps deployments with Docker and Kubernetes integration.
TensorRT Inference Server:
- Maximizes utilization by enabling inference for multiple models on one or more GPUs
- Supports all popular AI frameworks
- Supports audio streaming inputs
- Dynamically batches requests to increase throughput
- Provides latency and health metrics for auto scaling and load balancing
TensorRT Inference Server is available in a ready-to-deploy container from the NGC container registry, making it simple to use in production environments. It is also available as open source, allowing developers to customize and extend functionality of the software to fit their specific data center workflows.
With TensorRT Inference Server, there’s now a common solution for AI inference, allowing researchers to focus on creating high-quality trained models, DevOps engineers to focus on deployment, and developers to focus on their applications, without needing to reinvent the plumbing for each AI-powered application.
Learn more in the NVIDIA TensorRT Inference Server developer blog.
Download TensorRT Inference Server from the NGC container registry or access the open source project from GitHub.
(Click to Zoom)
TensorRT 7: What's New
TensorRT 7 provides capabilities to accelerate the complete real-time conversational AI pipeline by over 10x versus CPU with support for speech recognition, natural language understanding as well as text to speech models. The new version includes a new compiler and expanded support for the ONNX format.
Highlights
- New compiler to accelerate recurrent neural networks commonly used in speech and anomaly detection
- Support for over 20 new ONNX operations to accelerate key speech models such as BERT, TacoTron 2 and WaveRNN
- Expanded support for Dynamic Shapes to enable key conversational AI models
While TensorRT accelerates every framework, it is integrated with TensorFlow (v2.0 and v1.15), and ONNX Runtime 1.0 so that you can use TensorRT from within frameworks.
TensorRT 7 GA is available for download to members of the NVIDIA Developer Program. The latest version of plugins, parsers and new samples for BERT, Mask-RCNN, Faster-RCNN, NCF and OpenNMT are also available as open source from the TensorRT GitHub repository.
NVIDIA TensorRT Inference Server now supports ONNX graph and PyTorch backends, Model Control API for dynamic model loading/unloading, It is available as a ready-to-deploy container from NGC and as an open source project from GitHub.
Additional Resources
- TensorRT 4 Overview (Blog)
- “Hello World” For TensorRT(Sample Code)
- “Hello World” For TensorRT From ONNX(Sample Code)
- Performing Inference In INT8 Using Custom Calibration(Sample Code)
- Intro to RNNs in TensorRT (Video)
- Introduction to TensorRT (Webinar)
- 8-Bit Inference with TensorRT (Webinar)
- Real-Time Natural Language Understanding with BERT Using TensorRT (Blog)
- Introduction PART [1, 2, 3] (Blog)
- PyTorch - RNN (Blog)
- Neural Machine Translation (NMT) Using A Sequence To Sequence (seq2seq) Model(Sample Code)
- Building An RNN Network Layer By Layer(Sample Code)
- Optimizing` NMT with TensorRT (Webinar)
- Movie Recommendation Using Neural Collaborative Filter (NCF)(Sample Code)
- Deep Recommender (Sample Code)
- Intro to Recommenders in TensorRT (Video)
- Real-time object detection on GPUs in 10 mins (Blog)
- How to perform inference for common applications (Webinar)
- Creating object detection pipelines on GPUs (Blog)
- Object detection with SSD network (Python Code Sample)
- Object detection with SSD, Faster R-CNN networks (C++ Code Samples)
You can find additional resources on https://devblogs.nvidia.com/tag/tensorrt/ and interact with the TensorRT developer community on the TensorRT Forum
Get Started With Hands-On Training
The NVIDIA Deep Learning Institute (DLI) offers hands-on training for developers, data scientists, and researchers in AI and accelerated computing.Get hands-on experience with TensorRT in self-paced electives on Optimization and Deployment of TensorFlow Models with TensorRT and Deployment for Intelligent Video Analytics using TensorRT today.
Availability
TensorRT is freely available to members of the NVIDIA Developer Program from the TensorRT product page for development and deployment. The latest version of plugins, parsers and samples are also available as open source from the TensorRT github repository.
Developers can also get TensorRT in the TensorRT Container from the NGC container registry.
TensorRT Inference Server is available as a ready-to-deploy container from the NGC container registry and as an open source project from GitHub.
TensorRT is included in:
- NVIDIA Deepstream SDK for real-time streaming analytics in Computer Vision and Intelligent Video Analytics (IVA) applications
- NVIDIA DriveInstall for NVIDIA DRIVE PX2 autonomous driving platform
- NVIDIA Jetpack for Jetson TX1, TX2 embedded platforms