Revision History
Abstract
NVIDIA TensorRT is a C++ library that facilitates high performance inference on NVIDIA GPUs. It is designed to work in connection with deep learning frameworks that are commonly used for training. TensorRT focuses specifically on running an already trained network quickly and efficiently on a GPU for the purpose of generating a result; also known as inferencing. These release notes describe the key features, software enhancements and improvements, and known issues for the TensorRT 8.6.12 product package.
1. TensorRT for DRIVE OS
1.1. DRIVE OS Linux "Standard"
1.2. DRIVE OS QNX "Standard"
1.4. DRIVE OS for Safety Proxy
- The TensorRT proxy runtime is a version of the safety runtime for platforms that are not safety certified. This includes NVIDIA DRIVE OS x86 SDK, NVIDIA DRIVE OS Linux SDK, NVIDIA DRIVE OS Linux PDK, NVIDIA DRIVE OS QNX SDK and NVIDIA DRIVE OS QNX PDK. The proxy runtime is part of the development flow for safety but it is not certified itself. The proxy runtime only supports engines with engine capability kSAFETY (safe engines).
- Headers allow applications to compile against the proxy runtime and the safety runtime.
- The safety runtime is also a library that allows applications to load serialized engine plans and perform inference. It is only available for QNX safety. The safety runtime only supports engines with engine capability kSAFETY (safe engines).
3. New Features and Enhancements
ILayer Scope Relaxation
- IActivationLayer: the minimum rank of the input and output tensors for IActivationLayer is relaxed to 0.
- IConstantLayer: the batch, channel, and spatial dimension restrictions for IConstantLayer was removed.
- IElementWiseLayer: the minimum rank of the input and output tensors for IElementWiseLayer is relaxed to 0.
- IGatherLayer: the minimum rank of the input index tensors and output tensors for IGatherLayer is relaxed to 0. The minimum rank of the input data tensors for IGatherLayer is relaxed to 1.
- IIdentityLayer: the minimum rank of the input and output tensors for IIdentityLayer is relaxed to 0.
- IPluginV2Layer: the minimum rank of the input and output tensors for IPluginV2Layer is relaxed to 0. Index tensors with precision INT32 are supported.
- IScaleLayer: the batch, channel, and spatial dimension restrictions for IScaleLayer was removed.
- IShuffleLayer: the minimum rank of the output tensor for IShuffleLayer is relaxed to 0.
DLA Layer Scope Relaxation
- IReduceLayer: IReduceLayer with the Max operation is supported where any combination of the CHW is reduced.
- IElementWiseLayer: IElementWiseLayer with the Equal, Div, Pow, Greater, and Less operations is supported. The broadcasting is supported with restrictions.
- IUnaryLayer: IUnaryLayer operations: Sin, Cos, and Atan are supported.
TensorRT Standard Build
The TensorRT 8.6 release includes changes to the TensorRT 8.6.1 standard builder and runtime that appear in TensorRT for DRIVE OS 6.0. For more information, refer to the NVIDIA TensorRT 8.6.1 Release Notes.
Documentation Changes
- The NVIDIA TensorRT 8.6.12 Developer Guide for DRIVE OS is based on the enterprise TensorRT 8.6.1 release. We have modified the TensorRT 8.6.1 Developer Guide documentation for DRIVE OS 6.0.9 accuracy. The TensorRT safety content has been removed.
- The TensorRT safety content is in the NVIDIA TensorRT 8.6.12 Safety Developer Guide Supplement for DRIVE OS. Refer to this PDF for all TensorRT safety specific documentation.
4. Fixed Issues
| Reference ID | Module | Description |
|---|---|---|
| 4138970 | Consistency checker | The consistency checker will report an error when a standalone IScaleLayer is not fused with other layers and the TensorRT builder selects INT8 formats for its I/O tensor |
| 4157177 | TensorRT builder | An assertion error will occur from the TensorRT builder if an IActivationLayer serves as the input for an IElementwiseLayer and is also the input for another layer. |
| 4220077 | TensorRT builder | If the Matmul operator is followed by an Add operator, and the product of the input tensor dimensions (excluding the last dimension) of the Matmul operator is larger than 4096, the TensorRT builder may report an error. |
| 4125845 | TensorRT builder | Some networks with Convolution layers may fail to build when the builderOptimizationLevel is set to 4 or 5. |
| 4143531 | TensorRT runtime |
Networks may fail with safe CUDA error when switching to the Operational State for NvDVMS (DRIVE OS VM State Management) due to the violation to DOS_RES_107. This issue is fixed except under narrow circumstances. Refer to 4350817 under Known Issues for more information. |
5. Known Limitations
6. Known Issues
| Reference ID | Module | Description |
|---|---|---|
| 3656116 | TensorRT runtime |
What is the issue? There is an up to 7% performance regression for the 3D-UNet networks compared to TensorRT 8.4 EA when running in INT8 precision on NVIDIA Orin due to a functionality fix. How does it impact the customer? When running 3D-UNet networks in INT8 precision, the latency will be up to 7% longer than in TensorRT 8.4 EA. If there is a workaround, what is it? To work around this issue, set the input type and format to kINT8 and kCHW32, respectively. When can we expect the fix? We do not plan to fix this performance regression since it was caused by a necessary fix for an accuracy issue. Is it for Standard/Safety, SDK/PDK? Standard, SDK |
| 3263411 | TensorRT builder |
What is the issue? For some networks, building and running an engine in the standard runtime will have better performance than the safety runtime. This can be due to various limitations in scope of the safety runtime including more limited tactics, tensor size limits, and operations supported in the safety scope. How does it impact the customer? Inference in the safety runtime may be significantly slower than in the standard runtime. If there is a workaround, what is it? Depending on the network, it may or may not be possible to reorganize operations into a more efficient form matching the safety runtime scope. What is the recommendation? It is recommended to work with NVIDIA and provide proxy networks as early as possible that demonstrate key performance metrics close to actual production networks. Is it for Standard/Safety, SDK/PDK? Safety, SDK |
| 3793130 | TensorRT runtime |
What is the issue? Enabling the CUDA-graph option may cause the safety runtime to perform less efficiently compared to the proxy runtime for some networks. This discrepancy is due to the different objectives of the safety and proxy runtime. The safety runtime has more restrictive constraints to fulfill safety goals, resulting in different implementations between safety and proxy runtime. How does it impact the customer? Using the CUDA-graph for inference in the safety runtime may result in slower performance compared to the proxy runtime. However, this can vary depending on the inference network. If there is a workaround, what is it? It is recommended to check whether enabling CUDA-graph improves performance on the networks in production. Since the safety implementation with CUDA-graph comes with additional error checking and more deterministic execution, it is recommended to conduct cost-benefit analysis to decide if using CUDA-graph is beneficial to the use case. It is also recommended to work with NVIDIA and provide proxy networks as early as possible that demonstrate key performance metrics close to actual production networks. When can we expect the fix? In order to achieve safety, the implementation might require further support on error-checking and robustness measures. This could demand extra CPU/GPU cycles. However, in certain scenarios, the safety implementation might be faster since it does not support some features in proxy runtime. The performance parity will continue to improve in the future releases but it might not be completely realized. Is it for Standard/Safety, SDK/PDK? Safety, SDK |
| 4323665 | TensorRT runtime |
What is the issue? The TensorRT safety runtime does not set the CUDA API mode by invoking the API cudaSafeEXSelectAPIMode() at the initialization state. How does it impact the customer? The initialization stage of the TensorRT safety runtime does not prevent ASIL-D applications from using the TensorRT safety runtime. When this is fixed in the next release, applications that set the CUDA API mode as ASIL-D will not be able to use the TensorRT safety runtime. When can we expect the fix? The bug will be fixed by the next release. Is it for Standard/Safety, SDK/PDK? Safety, SDK |
| 4350817 | TensorRT runtime |
What is the issue? Some networks containing the softmax layer may fail with SafeCaskError when switching to the Operational State using NvDVMS (DRIVE OS VM State Management), which helps to verify DOS_RES_107. How does it impact the customer? For some networks containing softmax, the customer may not be able to switch to the Operational State using NvDVMS. If there is a workaround, what is it? Refrain from using NvDVMS for networks that contain the softmax layer. When can we expect the fix? The bug will be fixed by the next release. Is it for Standard/Safety, SDK/PDK? Safety, SDK |
| 4369190 | TensorRT safety samples |
What is the issue? The sample sampleSafeINT8 cannot run on the safety runtime due to a bug in the inference phase. How does it impact the customer? The customer cannot run sampleSafeINT8 on the safety runtime, only the proxy runtime. When can we expect the fix? This will be fixed in the next TensorRT release. Is it for Standard/Safety, SDK/PDK? Safety, SDK |
7. TensorRT Release Properties
| Linux x86-64 | Linux AArch64 | QNX AArch64 | ||
|---|---|---|---|---|
| QNX Safety | QNX Standard | |||
| Supported NVIDIA CUDA® versions | 11.4.28 | 11.4.28 | 11.4.28 | 11.4.28 |
| Supported NVIDIA cuDNN versions | 8.9.2 | 8.9.2 | No | 8.9.2 |
| TensorRT Python API | Yes | Yes | No | No |
| NvOnnxParser | Yes | Yes | No | Yes |
7.1. Hardware Precision
| CUDA Compute Capability | Example Device | TF32 | FP32 | FP16 | INT8 | FP16 Tensor Cores | INT8 Tensor Cores | DLA |
|---|---|---|---|---|---|---|---|---|
| 8.7 | NVIDIA Orin |
No (TensorRT safe) Yes (TensorRT standard) |
Yes | Yes | Yes | Yes | Yes | Yes |
| 8.6 | NVIDIA A10 | Yes | Yes | Yes | Yes | Yes | Yes | No |
| 8.0 | NVIDIA PG199 | Yes | Yes | Yes | Yes | Yes | Yes | No |
7.2. Software Versions Per Platform
| Platform | Compiler Version | Python Version |
|---|---|---|
| Ubuntu 20.04 x86-64 | gcc 9.3.0 | 3.8 |
| Ubuntu 20.04 AArch64 | gcc 9.3.0 | 3.8 |
| QNX AArch64 | QNX 7.1.0 Q++ 8.3.0 | N/A |
7.3. Compatibility
- CUDA 11.4.28
- cuDNN 8.9.2
- TensorFlow 1.15.5
- PyTorch 1.13.1
- ONNX 1.12.0 and opset 17
- DLA 3.14
Notice
Notice
This document is provided for information purposes only and shall not be regarded as a warranty of a certain functionality, condition, or quality of a product. NVIDIA Corporation (“NVIDIA”) makes no representations or warranties, expressed or implied, as to the accuracy or completeness of the information contained in this document and assumes no responsibility for any errors contained herein. NVIDIA shall have no liability for the consequences or use of such information or for any infringement of patents or other rights of third parties that may result from its use. This document is not a commitment to develop, release, or deliver any Material (defined below), code, or functionality.
NVIDIA reserves the right to make corrections, modifications, enhancements, improvements, and any other changes to this document, at any time without notice.
Customer should obtain the latest relevant information before placing orders and should verify that such information is current and complete.
NVIDIA products are sold subject to the NVIDIA standard terms and conditions of sale supplied at the time of order acknowledgement, unless otherwise agreed in an individual sales agreement signed by authorized representatives of NVIDIA and customer (“Terms of Sale”). NVIDIA hereby expressly objects to applying any customer general terms and conditions with regards to the purchase of the NVIDIA product referenced in this document. No contractual obligations are formed either directly or indirectly by this document.
NVIDIA products are not designed, authorized, or warranted to be suitable for use in medical, military, aircraft, space, or life support equipment, nor in applications where failure or malfunction of the NVIDIA product can reasonably be expected to result in personal injury, death, or property or environmental damage. NVIDIA accepts no liability for inclusion and/or use of NVIDIA products in such equipment or applications and therefore such inclusion and/or use is at customer’s own risk.
NVIDIA makes no representation or warranty that products based on this document will be suitable for any specified use. Testing of all parameters of each product is not necessarily performed by NVIDIA. It is customer’s sole responsibility to evaluate and determine the applicability of any information contained in this document, ensure the product is suitable and fit for the application planned by customer, and perform the necessary testing for the application in order to avoid a default of the application or the product. Weaknesses in customer’s product designs may affect the quality and reliability of the NVIDIA product and may result in additional or different conditions and/or requirements beyond those contained in this document. NVIDIA accepts no liability related to any default, damage, costs, or problem which may be based on or attributable to: (i) the use of the NVIDIA product in any manner that is contrary to this document or (ii) customer product designs.
No license, either expressed or implied, is granted under any NVIDIA patent right, copyright, or other NVIDIA intellectual property right under this document. Information published by NVIDIA regarding third-party products or services does not constitute a license from NVIDIA to use such products or services or a warranty or endorsement thereof. Use of such information may require a license from a third party under the patents or other intellectual property rights of the third party, or a license from NVIDIA under the patents or other intellectual property rights of NVIDIA.
Reproduction of information in this document is permissible only if approved in advance by NVIDIA in writing, reproduced without alteration and in full compliance with all applicable export laws and regulations, and accompanied by all associated conditions, limitations, and notices.
THIS DOCUMENT AND ALL NVIDIA DESIGN SPECIFICATIONS, REFERENCE BOARDS, FILES, DRAWINGS, DIAGNOSTICS, LISTS, AND OTHER DOCUMENTS (TOGETHER AND SEPARATELY, “MATERIALS”) ARE BEING PROVIDED “AS IS.” NVIDIA MAKES NO WARRANTIES, EXPRESSED, IMPLIED, STATUTORY, OR OTHERWISE WITH RESPECT TO THE MATERIALS, AND EXPRESSLY DISCLAIMS ALL IMPLIED WARRANTIES OF NONINFRINGEMENT, MERCHANTABILITY, AND FITNESS FOR A PARTICULAR PURPOSE. TO THE EXTENT NOT PROHIBITED BY LAW, IN NO EVENT WILL NVIDIA BE LIABLE FOR ANY DAMAGES, INCLUDING WITHOUT LIMITATION ANY DIRECT, INDIRECT, SPECIAL, INCIDENTAL, PUNITIVE, OR CONSEQUENTIAL DAMAGES, HOWEVER CAUSED AND REGARDLESS OF THE THEORY OF LIABILITY, ARISING OUT OF ANY USE OF THIS DOCUMENT, EVEN IF NVIDIA HAS BEEN ADVISED OF THE POSSIBILITY OF SUCH DAMAGES. Notwithstanding any damages that customer might incur for any reason whatsoever, NVIDIA’s aggregate and cumulative liability towards customer for the products described herein shall be limited in accordance with the Terms of Sale for the product.
Arm
Arm, AMBA and Arm Powered are registered trademarks of Arm Limited. Cortex, MPCore and Mali are trademarks of Arm Limited. "Arm" is used to represent Arm Holdings plc; its operating company Arm Limited; and the regional subsidiaries Arm Inc.; Arm KK; Arm Korea Limited.; Arm Taiwan Limited; Arm France SAS; Arm Consulting (Shanghai) Co. Ltd.; Arm Germany GmbH; Arm Embedded Technologies Pvt. Ltd.; Arm Norway, AS and Arm Sweden AB.
HDMI
HDMI, the HDMI logo, and High-Definition Multimedia Interface are trademarks or registered trademarks of HDMI Licensing LLC.
Blackberry/QNX
Copyright © 2020 BlackBerry Limited. All rights reserved.
Trademarks, including but not limited to BLACKBERRY, EMBLEM Design, QNX, AVIAGE, MOMENTICS, NEUTRINO and QNX CAR are the trademarks or registered trademarks of BlackBerry Limited, used under license, and the exclusive rights to such trademarks are expressly reserved.
Trademarks
NVIDIA, the NVIDIA logo, and CUDA, DALI, DGX-1, DRIVE, JetPack, Orin, Pegasus, TensorRT, Triton and Xavier are trademarks and/or registered trademarks of NVIDIA Corporation in the United States and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.