Version 450.51.05(Linux)/451.48(Windows)

Release notes for the Release 450 family of NVIDIA® Data Center GPU Drivers for Linux and Windows.

This edition of Release Notes describes the Release 450 family of NVIDIA® Data Center GPU Drivers for Linux and Windows. NVIDIA provides these notes to describe performance improvements, bug fixes and limitations in each documented version of the driver.

1. Version Highlights

This section provides highlights of the NVIDIA Data Center GPU R450 Driver (version 450.51.05 Linux and 451.48 Windows).

For changes related to the 450 release of the NVIDIA display driver, review the file "NVIDIA_Changelog" available in the .run installer packages.

Driver release date: 07/07/2020

1.1. Software Versions

07/14/2020: For this release, the software versions are listed below.

  • CUDA Toolkit 11: 11.0.194 (cudadrt)

    Note that starting with CUDA 11, individual components of the toolkit are versioned indepdently. For a full list of the individual versioned components (e.g. nvcc, CUDA libraries etc.), see the CUDA Toolkit Release Notes

  • NVIDIA Data Center GPU Driver: 450.51.05 (Linux) / 451.48 (Windows)

  • GPU VBIOS:

    • 92.00.19.00.10 (NVIDIA A100 SKU200 with heatsink for HGX A100 8-way and 4-way)

    • 92.00.19.00.13 (NVIDIA A100 SKU202 w/o heatsink for HGX A100 4-way)

  • The following software versions apply for HGX A100 systems.

    • Fabric Manager: 450.51.05 (Use nv-fabricmanager -v)

    • NVSwitch VBIOS: 92.10.14.00.01

    • NVFlash: 5.648

    • PEX88000 Firmware: 0.Ff_0

    • FPGA Firmware: 2.94

    • CEC: 3.5

    • PEX8725: 1.3

    • FRU EEPROM: 1.3

Due to a revision lock between the VBIOS and driver, VBIOS versions >= 92.00.18.00.00 must use corresponding drivers >= 450.36.01. Older VBIOS versions will work with newer drivers.

1.2. New Features

General

  • Added support for CUDA 11.0. For more information on CUDA 11.0, refer to the CUDA Toolkit 11.0 Release Notes

  • Added support for NVIDIA A100. The supported products are:

    • NVIDIA A100-SXM4-40GB
    • NVIDIA A100-PG509-200
    • NVIDIA A100-PCIE-40GB

  • Added support for Multi-Instance GPU (MIG) on A100. See the NVML API documentation and nvidia-smi mig command for more information on configuration and management of MIG instances. The following capabilities for MIG are supported for this release.

    • Support for running containers in MIG instances.

    • Support for nvJPEG.

    • Support for Video SDK (encode and decode).

    • Support for assignment of JPEG decoders and NVENC hardware units to MIG instances based on supported profiles.

    • Support for CUDA developer tools - debug and trace is supported for this release.

NVIDIA Management Library (NVML)

  • NVML Updates (Refer to the R450 NVML API Documentation for more information about these APIs)

    • New error code NVML_ERROR_INSUFFICIENT_RESOURCES when the GPU runs out critical resources, other than memory.

    • Added new counters for NVLink, including flit error counters, CRC data errors, replay errors, bandwidth, speed, throughput and row remapper. See the NVML API documentation for more information on these new counters.

    • Added support for MIG configuration and management. See the NVML APIs for setting up MIG mode, managing GPU and Compute instances.

    • Added nvmlDeviceGetRemappedRows to return the number of rows remapped by the A100 GPU.

    • The following NVLink APIs are deprecated.

      • nvmlDeviceSetNvLinkUtilizationControl()

      • nvmlDeviceGetNvLinkUtilizationControl()

      • nvmlDeviceGetNvLinkUtilizationCounter()

      • nvmlDeviceFreezeNvLinkUtilizationCounter()

      • nvmlDeviceResetNvLinkUtilizationCounter()

    • nvmlDeviceGetNvLinkUtilizationCounter() is now replaced by nvmlDeviceGetFieldValues(). See the NVML API documentation for the new field parameters.

Fabric Manager for NVSwitch Systems

  • Fabric Manager is now available as via standalone installer packages (.deb/.rpm) for Linux distributions (nvidia-fabricmanager-driver-version)

  • Added the ability for CUDA workloads to continue running even when the Fabric Manager is stopped or terminated.

  • Added a new API fmGetUnsupportedFabricPartitions() to list unsupported/degraded partitions. The API reports all the GPUs in each unsupported partition by physical ID.

  • Added support for in-place package upgrades to the Fabric Manager (without updating the driver or) without affecting running workloads.

  • APIs to query the number of NVLinks and the available bandwidth between these links.

  • Added ability to support 4 GPUs per HGX baseboard (to create an 8x partition size).

  • FM has a default NVLink partition grouping information and provides options for overriding the defaults with custom partition information.

  • The nvswitch-audit tool is now included in the Fabric Manager .deb and .rpm packages.

  • Added configuration options for Fabric Manager to prevent termination due to a fatal NVLINK error, NVSwitch error, or other typical software events. In the case of a fatal error, CUDA application creation will be blocked and any running application will be terminated.

Systems

  • Added the ability for NVLink to switch to single lane mode to save power at idle.

  • Added support for multi-host configurations. Note that this is supported only in 16-GPU systems (with two HGX A100 baseboards).

  • Added support for GPU high availability mode - where failures of up to 1 GPU is supported on HGX A100. In the shared NVSwitch scenario, Fabric Manager disables partitions with a failing GPU and reports availability of all other partitions. Refer to the FM User Guide section on GPU Failure for more information.

  • Added support for NVSwitch degraded mode - where the system remains functional but peer NVSwitch trunk links will operate at reduced bandwidth. Fabric Manager is also configurable to abort on NVSwitch failure. Refer to the FM User Guide for more information.

  • HGX A100 now automatically configures the droopy value for A100 SXM4 modules based on the input voltage. Note that there is no user facing API or knob for controlling this behavior.

  • On a GPU reset, the hardware now remaps rows that reported errors (single-bit or double-bit) to new memory locations to preserve contiguous memory. See the application note "NVIDIA A100 GPU Memory Error Management Application Note" for more information on this capability.

  • Added support for systems that include 2x 8 GPU baseboards (up to 16 GPUs)

  • Resilient errors (in XIDs) are isolated to the process and failing CUDA applications are terminated but other processes on the GPU or partition can run to completion (or new processes can be launched) without requiring a GPU reset.

SMBPBI for NVIDIA Data Center GPUs

  • The following SMBPBI APIs were added for out-of-band management of the GPU and NVSwitches.

    • API to read the default value and range of possible max settings for the application and memory clocks.

    • API to read the current value of the application and memory clocks.

    • API for reporting GPU power cap values.

    • API for changing power cap configurations.

    • API for reporting the state of NVSwitch state (active or degraded) along with the reason for the degraded state.

    • API for reporting current running state of Fabric Manager.

    • API for reporting inactive NVLinks.

    • API for reporting failures when performing row-remapping updates.

    • API for reporting row-remapping pending state.

    • API for reporting number of remapped rows.

    • API for reporting GPU XID errors.

    • API for reporting PCIe width and speed.

    • API for reporting PCIe link status.

    • API for reporting PCIe error count.

    • API for reporting Target temperature.

    • API for reporting Default TGP.

    • API for reporting Power Capping.

    • API for reporting NVSwitch NVLink reads and writes for both payload traffic and total traffic (including protocol overhead).

    • API to read HBM temperature.

    • APIs to query to report the NVLink error state even if the NVLink in error is set to 'OFF'.

    • API for reporting GPU NVLink reads and writes for both payload traffic and total traffic (including protocol overhead).

    • The following MIG related features are supported. Note that the other out-of-band APIs for reporting metrics or management are at the granularity of GPUs rather than instances when the GPU is in MIG mode.

      • API to toggle MIG mode on GPUs.

      • API for reporting number of active GPU instances when the GPU is in MIG mode.

    • API for reporting XID errors via polling.

    • Ability to disable a GPU from the system.

    • Ability to set clock limits on the GPUs.

    • Ability to disable an entire tray even with the NVLINK bridge connected.

  • Added a number of performance improvements for out-of-band management using SMBPBI

    • Merged SMBPBI status and data into one request with 3-bytes (24 bits) of data per read

    • Bundled GPU telemetry using user selectable commands (such as GPU power, GPU temperature, HBM temperature) thereby reducing the number of reads required

    • Increased data rate over SMBUS from 100KHz to 400KHz

1.3. Fixed Issues

  • Various security issues were addressed, for additional details on the med-high severity issues please review NVIDIA Product Security for more information.

  • Fixed an issue where NVML/nvidia-smi would report up to 25% utilization of the GPU when idle.

  • Fixed an issue in shared NVSwitch virtualization mode where SXid errors may not be generated when there is a PCIe reset to GPUs in an activated partition.

  • Fixed an issue where the NVIDIA kernel driver may fail to query the GPU's UUID if the query is attempted right after GPU reset.

  • Fixed an issue where reseting a single GPU (with active NVLinks) in a multi-GPU NVSwitch system would result in the Fabric Manager service to hang after a restart of the service

1.4. Known Issues

General

  • By default, Fabric Manager runs as a systemd service. If using DAEMONIZE=0 in the Fabric Manager configuration file, then the following steps may be required.

    1. Disable FM service from auto starting. (systemctl disable nvidia-fabricmanager)

    2. Once the system is booted, manually start FM process. (/usr/bin/nv-fabricmanager -c /usr/share/nvidia/nvswitch/fabricmanager.cfg). Note, since the process is not a daemon, the SSH/Shell prompt will not be returned (use another SSH shell for other activities or run FM as a background task).

  • On A100-PCIE-40GB, when two GPUs are connected using NVlink Bridge, there is a less than 5% performance degradation on tested workloads. This issue will be addressed in future driver release.

  • When starting the Fabric Manager service, the following error may be reported: detected NVSwitch non-fatal error 10003 on NVSwitch pci. This error is not fatal and no functionality is affected. This issue will be resolved in a future driver release.

  • On NVSwitch systems with Windows Server 2019 in shared NVSwitch virtualization mode, the host may hang or crash when a GPU is disabled in the guest VM. This issue is under investigation.

  • In some cases, after a system reboot, the first run of nvidia-smi shows an ERR! for the power status of a GPU in a multi-GPU A100 system. This issue is not observed when running with peristence mode.

GPU Performance Counters

The use of developer tools from NVIDIA that access various performance counters requires administrator privileges. See this note for more details. For example, reading NVLink utilization metrics from nvidia-smi (nvidia-smi nvlink -g 0) would require administrator privileges.

NoScanout Mode

NoScanout mode is no longer supported on NVIDIA Tesla products. If NoScanout mode was previously used, then the following line in the “screen” section of /etc/X11/xorg.conf should be removed to ensure that X server starts on Tesla products: 

Option         "UseDisplayDevice" "None"

Tesla products now support one display of up to 4K resolution.

Unified Memory Support

Some Unified Memory APIs (for example, CPU page faults) are not supported on Windows in this version of the driver. Review the CUDA Programming Guide on the system requirements for Unified Memory

CUDA and unified memory is not supported when used with Linux power management states S3/S4.

IMPU FRU for Volta GPUs

The driver does not support the IPMI FRU multi-record information structure for NVLink. See the Design Guide for Tesla P100 and Tesla V100-SXM2 for more information.

Video Memory Support

For Windows 7 64-bit, this driver recognizes up to the total available video memory on Tesla cards for Direct3D and OpenGL applications.

For Windows 7 32-bit, this driver recognizes only up to 4 GB of video memory on Tesla cards for DirectX, OpenGL, and CUDA applications.

Experimental OpenCL Features

Select features in OpenCL 2.0 are available in the driver for evaluation purposes only.

The following are the features as well as a description of known issues with these features in the driver:

Device side enqueue

  • The current implementation is limited to 64-bit platforms only.

  • OpenCL 2.0 allows kernels to be enqueued with global_work_size larger than the compute capability of the NVIDIA GPU. The current implementation supports only combinations of global_work_size and local_work_size that are within the compute capability of the NVIDIA GPU. The maximum supported CUDA grid and block size of NVIDIA GPUs is available at http://docs.nvidia.com/cuda/cuda-c-programming-guide/index.html#computecapabilities. For a given grid dimension, the global_work_size can be determined by CUDA grid size x CUDA block size.

  • For executing kernels (whether from the host or the device), OpenCL 2.0 supports non-uniform ND-ranges where global_work_size does not need to be divisible by the local_work_size. This capability is not yet supported in the NVIDIA driver, and therefore not supported for device side kernel enqueues.

Shared virtual memory

  • The current implementation of shared virtual memory is limited to 64-bit platforms only.

2. Virtualization

To make use of GPU passthrough with virtual machines running Windows and Linux, the hardware platform must support the following features:

  • A CPU with hardware-assisted instruction set virtualization: Intel VT-x or AMD-V.

  • Platform support for I/O DMA remapping.

  • On Intel platforms the DMA remapper technology is called Intel VT-d.

  • On AMD platforms it is called AMD IOMMU.

Support for these feature varies by processor family, product, and system, and should be verified at the manufacturer's website.

Supported Hypervisors

The following hypervisors are supported:

Hypervisor Notes
Citrix XenServer Version 6.0 and later
VMware vSphere (ESX / ESXi) Version 5.1 and later.
Red Hat KVM Red Hat Enterprise Linux 7 with KVM
Microsoft Hyper-V

Windows Server 2016 Hyper-V Generation 2
Tesla products now support one display of up to 4K resolution.

Supported Graphics Cards

The following GPUs are supported for device passthrough:

GPU Family Boards Supported
NVIDIA Ampere GPU Architecture NVIDIA A100
Turing NVIDIA T4
Volta NVIDIA V100
Pascal

Tesla: P100, P40, P4
Maxwell

Tesla: M60, M40, M6, M4
Kepler

Tesla: K520, K80

3. Hardware and Software Support

Support for these feature varies by processor family, product, and system, and should be verified at the manufacturer's website.

Supported Operating Systems for NVIDIA Data Center GPUs

The Release 450 driver is supported on the following operating systems:

  • Windows x86_64 operating systems:

    • Microsoft Windows® Server 2019

    • Microsoft Windows® Server 2016

    • Microsoft Windows® 10

  • The table below summarizes the supported Linux 64-bit distributions. For a complete list of distributions, kernel versions supported, see the CUDA Linux System Requirements documentation.

    Distribution x86_64 POWER Arm64 Server
    OpenSUSE Leap 15.1 Yes No No
    Red Hat Enterprise Linux / CentOS 8.y (where y <= 2) Yes Yes Yes
    Red Hat Enterprise Linux / CentOS 7.y (where y <= 8) Yes No No
    SUSE Linux Enterprise Server 15.1 Yes No Yes (see note)
    Ubuntu 20.04 LTS Yes No No
    Ubuntu 18.04.z LTS (where z <= 4) Yes Yes Yes
    Ubuntu 16.04.z LTS (where z <= 6) Yes No No

    Note that SUSE Linux Enterprise Server (SLES) 15.1 is provided as a preview for Arm64 server since there are known issues when running some CUDA applications related to dependencies on glibc 2.27.

Supported Operating Systems and CPU Configurations for HGX A100

The Release 450 driver is validated with HGX A100 on the following operating systems and CPU configurations:

  • Linux 64-bit distributions:

    • Red Hat Enterprise Linux 8.1 (in 4/8/16-GPU configurations)

    • CentOS Linux 7.7 (in 4/8/16-GPU configurations)

    • Ubuntu 18.04.4 LTS (in 4/8/16-GPU configurations)

    • SUSE SLES 15.1 (in 4/8/16-GPU configurations)

  • Windows 64-bit distributions:

    • Windows Server 2019 (in 4/8/16-GPU configurations)

  • CPU Configurations:

    • AMD Rome in PCIe Gen4 mode

    • Intel Skylake/Cascade Lake (4-socket) in PCIe Gen3 mode

Supported Virtualization Configurations

The Release 450 driver is validated with HGX A100 on the following configurations:

  • Passthrough (full visibility of GPUs and NVSwitches to guest VMs):

    • 8-GPU configurations with Ubuntu 18.04.4 LTS

  • Shared NVSwitch (guest VMs only have visibility of GPUs and full NVLink bandwidth between GPUs in the same guest VM):

    • 16-GPU configurations with Ubuntu 18.04.4 LTS

API Support

This release supports the following APIs:

  • NVIDIA® CUDA® 11.0 for NVIDIA® KeplerTM, MaxwellTM, PascalTM, VoltaTM, TuringTM and NVIDIA Ampere architecture GPUs

  • OpenGL® 4.5

  • Vulkan® 1.1

  • DirectX 11

  • DirectX 12 (Windows 10)

  • Open Computing Language (OpenCLTM software) 1.2

Note that for using graphics APIs on Windows (i.e. OpenGL, Vulkan, DirectX 11 and DirectX 12) or any WDDM 2.0+ based functionality on Tesla GPUs, vGPU is required. See the vGPU documentation for more information.

Supported NVIDIA Data Center GPUs

The NVIDIA Data Center GPU driver package is designed for systems that have one or more Tesla products installed. This release of the driver supports CUDA C/C++ applications and libraries that rely on the CUDA C Runtime and/or CUDA Driver API.

NVIDIA Server Platforms
Product Architecture
NVIDIA HGX A100 A100 and NVSwitch
NVIDIA HGX-2 V100 and NVSwitch
A-Series Products
Product GPU Architecture
NVIDIA A100 NVIDIA Ampere
T-Series Products
Product GPU Architecture
NVIDIA T4 Turing
V-Series Products
Product GPU Architecture
NVIDIA V100 Volta
Tesla P-Series Products
Product GPU Architecture
NVIDIA Tesla P100 Pascal
NVIDIA Tesla P40 Pascal
NVIDIA Tesla P4 Pascal
Tesla K-Series Products
Product GPU Architecture
NVIDIA Tesla K520 Kepler
NVIDIA Tesla K80 Kepler
Tesla M-Class Products
Product GPU Architecture
NVIDIA Tesla M60 Maxwell
NVIDIA Tesla M40 24 GB Maxwell
NVIDIA Tesla M40 Maxwell
NVIDIA Tesla M6 Maxwell
NVIDIA Tesla M4 Maxwell

Notices

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.

Trademarks

NVIDIA and the NVIDIA logo are trademarks and/or registered trademarks of NVIDIA Corporation in the Unites States and other countries. Other company and product names may be trademarks of the respective companies with which they are associated.