GRID SDK FAQ - NVIDIA Developer
Contents
1. XenServer with NMOS Q25 What is the recommended SBIOS IPMI Firmware I should use A25 Please contact your NVIDIA GRID support person for the recommended SBIOS IPMI firmware Q26 Are there any special setting I need to do in SBIOS and IPMI for using these Servers for GRID SDK usage A26 Yes For SBIOS Setting VT d for Intel platforms and IOMMU for AMD platforms has to be enabled for virtualization environment Note For IPMI the FAN speed must be set to optimal for Virtualization and Baremetal environments Q27 For AMD servers i e Super Micro 1022GG when obtaining the system topology information I am unable to determine which NUMA domain the GRID device is closest to This system is a NUMA system and the System BIOS for NUMA is enabled However it fails when making the function call to retrieve the NUMA information How do I resolve the problem GRID SDK FAQ PG 06183 001 10 A27 Make sure you apply the motherboard has the latest SBIOS H8DGG 926 or newer for the SuperMicro SM 1022 server Update it to this SBIOS if it is older than this one In order to verify that the GPU and CPU are on the same NUMA node use a bandwidth test application Nvidia CUDA SDK contains a test application named bandwidthTest exe that is suitable for this purpose to measure the PCI e bandwidth between the CPU and GPU You can compare other NUMA nodes to make sure you are getting the best2. sys 2027er trf cfm Supermicro 2027 GRTRF Intel Platform http www supermicro com products system 1u 1027 sys 1027 r trf cfm Supermicro 1022 GG TF AMD Platform http www supermicro com aplus system 1u 1022 as 1022 tf cfm Q24 What are the recommended Board configurations in a Baremetal system A24 Servers from SuperMicro and Asus are recommended The actual number and Boards that should be used depends on the server type and use case The following tables give information about the maximum number of GPUs that can be supported Note For details on how to setup a system with 5xK340 we will be providing this soon Supermicro SMCServer 2027 Max K340 Max GK107 Max K520 Max GK104 Intel boards GPUs boards GPUs Windows Baremetal 3xK340 12 5xK520 10 Linux Baremetal 5xK340 20 5xK520 10 Windows Linux VMs using 5xK340 20 5xK520 10 XenServer with NMOS GRID SDK FAQ PG 06183 001 9 Supermicro SMC Server 1027 Max K340 Max GK107 Max K520 Max GK104 Intel boards GPUs boards GPUs Windows Baremetal 3xK340 12 3xK520 6 Linux Baremetal 3xK340 12 3xK520 6 Windows Linux VMs using 3xK340 12 3xK520 6 XenServer with NMOS Supermicro SMC 1022 AMD Max K340 Max GK107 Max K520 Max GK104 boards GPUs boards GPUs Windows Baremetal 2xK340 8 2xK520 4 Linux Baremetal 2xK340 8 2xK520 4 Windows Linux VMs using 2xK340 8 2xK520 4
3. EnableIntraRefresh 1 during encoder setup and then set bStartIntraRefresh 1 and dwIntraRefreshCnt n where n is the number of slices in which intra refresh should be split In other words the GRID SDK FAQ PG 06183 001 12 next n frames will be split in approximately equal number of sections on an MB row boundary and starting from top each section will be refreshed with intra MB s At the end of n frames the entire picture will be refreshed This method has the advantage of lower channel overhead since only part of the frame needs to intra coded but the disadvantage that it takes slightly longer to recover from the error e The third method for error recovery is to use reference picture invalidation This method works only if there is an upstream communication channel from the decoder to the encoder To use this method the decoder signals a lost frame to the encoder and the encoder then invalidates the reference frame by setting bInvalidateRefrenceFrames 1 and indicate the frames to be invalidated by setting dwNumRefFramesToInvalidate and corresponding timestamps in for frames to be invalidated in ulInvalidFrameTimeStamp NVFBC_MAX REF FRAMES For this to work the encoder must provide a monotonically increasing timestamp to each frame being encoded through the parameter ulCaptureTimeStamp The reference picture invalidation logic simply matches this timestamp with those passed in ulInvalidFrameTimeSta
4. GridPublicSupport nvidia com Q12 What are the differences between the GRID cloud gaming boards A12 GRID K520 and K340 are for use to stream consumer applications games GRID K520 which includes two Kepler based GPUs and 8GB of memory is designed to be a high performance GPU with a powerful 3D engine but less simultaneous encoding performance as GRID K340 NVIDIA recommends using the GRID K520 with games that require more graphics performance For more concurrent sessions with applications that require less graphics performance NVIDIA recommends the GRID K340 which includes four Kepler based GPUs and 4GB of memory The GRID K2 and K1 are for use to enable rich graphics in virtualized environments GRID K2 is designed for GRID K2 boards which include two higher end Kepler GPUs and 8GB of memory deliver maximum density for users of graphics intensive applications GRID K1 boards which include four Kepler based GPUs and 16GB of memory are designed to host the maximum number of concurrent users 3D Performance NVENC Encoder Memory GB per GPU per GPU per GPU GRID SDK FAQ PG 06183 001 5 Q13 How many simultaneous Capture and HW encode sessions can run on a single GRID K340 or K1 board How about for a single GRID K520 or K2 board A13 Each Kepler GPU has one NVENC hardware encoder that is capable of supporting up to 6 H 264 high quality 720p at 30 fps With GRID K2 and K520 boards there are 2 Kepler based GPUs per board P
5. I O throughput using the bandwidthTest exe tool from the CUDA Toolkit Q28 When setting up a new Server in a Baremetal environment there are problems installing and using the GRID devices The Server OS is Windows 7 or Windows 2008 R2 The NVIDIA driver installation succeeds but there are problems recognizing the GRID GPUs when I launch my application What is the solution A28 Please refer the GRID Game Server Configuration document that is included with the SDK Sections 1 2 explain how to configure the server and setup the driver Before installing the NVIDIA driver please make sure a You are connected through VNC or TeamViewer Microsoft Remote Desktop uses a software VGA driver which does not enable the NVIDIA GPU driver b Onboard VGA needs to be disabled Prior to installing the NVIDIA drivers the onboard VGA must be disabled Refer to the SuperMicro server manual or the GRID Game Server Configuration Guide for information how to disable VGA jumper JPG1 For Windows 7 and Windows Server 2008 R2 two GPUs from different vendors cannot be used as this is a limitation of the OS and one GPU vendor driver can be installed and running at atime If the onboard VGA with the NVIDIA GPUs is required with the GRID server Windows 8 and Windows Server 2012 both support this Note A benefit for using Windows 8 and Windows Server 2012 is that both allow IPMI management tools to run If the onboard VGA is disabled IPMI will is not avail
6. ITY This rate control mode allows higher bit budget for I frames and other high complexity frames e g scene changes resulting in better quality I frames Note however that both 2 pass modes NVFBC_H264 ENC PARAMS RC 2 PASS QUALITY and NVFBC_H264 ENC PARAMS RC_2 PASS FRAMESIZE_ CAP have lower performance than the single pass modes such as NVFBC_H264 ENC PARAMS RC_CBR If running in a streaming cloud gaming application set dwVBVBufferSize dwAvgBitRate dwFrameRateNum dwFrameRateDen dwVBVInitialDelay dwVBVBufferSize The above setting enables a special quality optimized low latency mode inside the hardware encoder Q31 When streaming an H 264 bitstream what are some of the methods I can use to recover from channel errors and lost frames on the decoder A31 If the decoder loses an entire frame you can use the following error recovery mechanisms Force an IDR frame to be generated by setting bForceIDRFrame 1 for the next frame on the encoder This is the easiest and brute force method for error recovery However this method is prone to problems because the IDR frame generated may have a lower quality larger size depending upon which RC mode is in use and since the channel is already bad there are higher chances of the IDR frame itself getting lost resulting in cyclically worsening problem Force an intra refresh wave This is the recommended method for error recovery To use this method set b
7. NVIDIA GRID SDK FAQ PG 06183 001 DOCUMENT CHANGE HISTORY Version Date Authors Description of Change 0 5 5 13 2013 BO Initial draft 0 8 12 5 2013 EY Revised and added more FAQ 0 9 12 5 2013 NS DG Added 6 new FAQ items 1 0 1 17 2014 EY First public release GRID SDK FAQ PG 06183 001 ii Q1 What is the GRID SDK How do developers get access to it A1 The GRID SDK enables fast capture and compression of the desktop display or render targets from NVIDIA GRID cloud gaming graphics boards You will need to register in order to download the GRID SDK Registration is free and the download page can be found here https developer nvidia com grid app game streaming Q2 I download the SDK Now how do I get started A2 For recommended hardware please refer to Question 23 of this FAQ Please refer to the Guide to Setup a GRID Server and the GRID SDK Programming Guide for how to begin using the SDK once you have downloaded the SDK and driver These samples are simple examples for how to use the GRID APIs on Linux or Windows Installation Directory Samples Q3 What components are included with GRID SDK What are they used for A3 The GRID SDK consists of two component software APIs NvFBC and NVIFR a NvFBC captures and optionally H 264 encodes the entire visible desktop b NvIFR captures and optionally H 264 encodes from a specific render target Note Both of these GRID
8. O 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 Information furnished is believed to be accurate and reliable However NVIDIA Corporation assumes no responsibility for the consequences of use of such information or for any infringement of patents or other rights of third parties that may result from its use No license is granted by implication of otherwise under any patent rights of NVIDIA Corporation Specifications mentioned in this publication are subject to change without notice This publication supersedes and replaces all other information previously supplied NVIDIA Corporation products are not authorized as critical components in life support devices or systems without express written approval of NVIDIA Corporation HDMI HDMI the HDMI logo and High Definition Multimedia Interface are trademarks or registered trademarks of HDMI Licensing LLC OpenCL OpenCL is a trademark of Apple Inc used under license to the Khronos Group Inc Trademarks NVIDIA and the NVIDIA logo are trademarks or registered trademarks of NVIDIA Corporation in the U S and other countries Other company and product names may be trademarks of the respective companies with which they are associated Copyright 2011 2014 NVIDIA Corporation All rights reserved GRID SDK FAQ PG 06183 001 14
9. able Q29 For a Baremetal server configuration when there are very many applications or game sessions using NvIFR are being launched on multiple GPUs the performance does not scale as expected A29 Included with the GRID SDK package there is a registry key modification Under the following folder and click on the MemoryManagerListSize_96MB reg file to install into your registry on the server This is required in order to achieve proper scaling across all of the GPUs Installation Directory Tools DXG_Kernel_Memory_Limit GRID SDK FAQ PG 06183 001 11 Q30 The quality of the H 264 bitstream obtained from the encoder does not appear to be good Are there settings that should be checked A30 The video encoder presets and RC modes exposed in the GRID SDK are optimized for use cases such as remote streaming cloud gaming remote desktop of applications etc There is no one set of parameters which will result in good quality under all conditions For general use cases please follow these guidelines Try setting the GOP length to 1 infinite If you set GOP length too small frequent I frames will be generated If the encoder is set up in low latency mode quality of I frames is low bit budget allocated for I frames is low resulting in frequent flickering effect If you must use periodic I frames i e if you cannot set GOP length 1 then try using the rate control mode eRateControl NVFBC_H264 ENC PARAMS RC 2 PASS QUAL
10. components will interface with the NVENC API for H 264 hardware compression The NVENC interface is not provided with the GRID SDK but NvFBC and NvIFR include functions to configure the H 264 hardware encoder Q4 What are the interfaces for NvFBC Capture A4 There is a NvFBC interface for each type of capture you want to do and where you want the data to go a NVFBC CAPTURE TO SYS captures the desktop and copies it to pinned system memory on the host b NVFBC_CAPTURE SHARED CUDA captures the desktop and copies it to CUDA device memory on the GPU c NVFBC CAPTURE TO H264 HW ENCODER captures the desktop compresses it using on chip hardware video encoder to a H 264 video stream and copies the H 264 encoded elementary bitstream to system memory on the host a www nvidia com NVI DIA Q5 What are the interfaces for NvIFR Capture A5 There is an NvIFR interface for each operation that you want to do when capturing a NvIFRToSys captures the render target and copies it to pinned system memory on the host b NvIFRH264 captures the render target compresses it using on chip hardware video encoder to a H 264 video stream and copies the H 264 encoded elementary bitstream to system memory on the host Q6 When should NvFBC and NvIFR be used A6 NvIER is the preferred solution to capture the video output of one specific application NvFBC is better for remote desktop applications For a more deta
11. erformance will more likely be hardware encoder limited if more sessions are running concurrently GRID K1 and K340 boards there are four Kepler based GPUs per board Performance will more likely be limited by the 3D graphic engine It does have the advantage of having twice as many encoders as K2 and K520 The GPUs in the GRID K520 and K2 are more powerful with higher fill rate pixels sec and geometry rates than the K340 and K1 able to generate more frames second than the NVENC encoder is able to compress The total number of concurrent sessions for these boards will depend on the games or applications being streamed the hardware encoder settings being used and the CPUs in the system Q14 The maximal intra refresh count is 16 How does NvIFR determine the number or size of macroblocks Will it be different if we encode with different resolutions Such 800x600 1280x720 A14 The count 16 is a limitation of GRID software and does not depend on resolution The refresh boundaries however will depend on the resolution The refresh regions always begin and end at the macroblock row boundary The boundaries are determined based on the resolution and the intra refresh count Q15 When intra refresh capability is enabled the frame will be divided into several slices Are there any issues with decoupling the intra refresh and slice capability A15 Each refresh is a slice because the GPU hardware allows changing motion vector search patterns in
12. iled explanation of the differences please refer to the Amazon G2 Instance Getting Started Guide included with the SDK Q7 Can NvIFR capture GDI 2D components of an application window If not will it be support in the future A7 No NvIER is the API for capturing from RenderTargets If you would like to capture GDI and 2D Components from the system desktop use NvFBC There are currently no plans to support this in the future Q8 What hardware and drivers are required for the GRID SDK A8 Please refer to the release notes included in the GRID SDK package you intend to use for supported hardware and operating systems Q9 What Operating Systems does the GRID SDK support A9 Microsoft Windows 7 Windows 8 Windows Server 2008 R2 Windows Server 2012 and Linux OS For servers running Windows we recommend Windows 8 and Windows Server 2012 because these operating systems are more efficient and can serve more concurrent users GRID SDK FAQ PG 06183 001 4 Q10 Can I distribute any GRID SDK components to my customers A10 No the GRID SDK is subject to NVIDIA s Software License Agreement which is included with the SDK Any re distribution requires express agreement from NVIDIA Q11 Whom can I contact on questions integrating my software with the GRID SDK A11 Please refer to the Programming Guide and included documentation If there are other more detailed questions please contact GRID developer support at NVIDIA
13. l of the NVIDIA GPU accelerated capabilities and does not enable the NVIDIA GRID driver Any applications running under Microsoft Remote Desktop will not be using the NVIDIA driver and will not have full benefits of GPU acceleration Q17 How do I perform DirectX 9 Capture and Encode using the GRID SDK A17 Please refer to the GRID SDK Sample Description document for guidelines for all of the GRID SDK samples For simple examples of how to use the API please refer to these four GRID SDK samples to get started NvFBCH264 DX9TFRSimpleH264HWEncode DX9IFRAsyncH264HWEncode DX 9IFRSharedSurfaceH264HWEncode Note there are other things to consider for your DirectX9 application that needs to be properly handled with software When the game loading is complete the D3DDevice becomes invalid resulting in a game hang IDirect3DDevice9 TestCooperativeLevel function will return D3D ERR_D EVICENOTRESET The Shim layer should add check before calling NvIFRTransferRenderTargetToH264HWEncoder whether D3D9Device is alive The application can call IDirect 3DDevice9 TestCooperativeLevel function and if it returns DJDERR_DEVICENOTRESET try calling IDirect3DDevice9 Reset function If a game instance creates a new device the shim layer will need to destroy the previous NvIFR context After that it can create new NvIFR context with the ne
14. mp and invalidates those reference frames This prevents the invalidated frames from being used for subsequently encoded frames as references and prevents error propagation When encoder exhausts all reference frames set via dwMaxNumRefFrames it automatically generates an IDR frame Note In addition to the above you can also use the error concealment on the decoder so as to avoid discarding an entire frame on the decoder This works if the error impact is not large Q32 How do I capture the mouse input while using NvFBC A32 The choice of having HW rendered mouse cursor blended on to the image grabbed by NvFBC is available only with the NvFBCToSys interface The other two interfaces always blend the HW rendered mouse cursor on the desktop at server end For NvFBCToSys to enable capturing desktop images with the HW cursor blended on to the grabbed image please refer to the documentation for the parameter NVFBC_TOSYS_SETUP_PARAMS bWithHW Cursor in the GRID SDK Programming Guide In order to enable capturing HW mouse images separately so that they can be blended at client side during streaming please refer to the documentation for the API NvFBCtoSysGrabMouse in the GRID SDK Programming Guide GRID SDK FAQ PG 06183 001 13 Notice ALL NVIDIA DESIGN SPECIFICATIONS REFERENCE BOARDS FILES DRAWINGS DIAGNOSTICS LISTS AND OTHER DOCUMENTS TOGETHER AND SEPARATELY MATERIALS ARE BEING PROVIDED AS IS NVIDIA MAKES N
15. nd performance Other hypervisors which are supported are Citrix VMware and KVM XCP 1 5 is now deprecated Note When using XenServer 6 2 with a K340 in a virtualized environment the Xen distribution needs to be patched to enable device class configuration space matching For K520 based servers this step is not required You can find the patch under the XenPatch folder in the SDK In the public SDK it includes the path for Xen For Xen to apply the patch gt rpm Uvh xen device model 1 8 0 89 7554 i686 rpm This is a required step to ensure that the Windows VM running on Xen can boot with GPU pass through GRID SDK FAQ PG 06183 001 8 Q21 Which Virtual Environments are supported by the GRID SDK A21 NMOS NVIDIA Multi OS One GPU is attached to one Virtual Machine Q22 Can hybrid card combinations of GRID boards be used in GRID servers A22 Using hybrid combinations is not recommended For example mixing K340 and K520 GPUs in the same system may work fine but is not the recommended way of setting up a baremetal or virtualized system Q23 Do I need a specific hardware system server to use GRID SDK A23 Following Servers and GPUs are recommended for use with the GRID SDK GRID GPUs NVIDIA GRID K340 K520 K1 and K2 Desktop GPUs NVIDIA Quadro K4000 and higher Mobile GPUs NVIDIA Quadro K2000M and higher GRID Servers Supermicro 1027 GRTRF Intel Platform http www supermicro com products system 2u 2027
16. tra vs inter only on slice boundary hence each refresh region has to be a slice Q16 Which Remote Desktop solution is recommended to connect with a GRID server or Amazon G2 instance Can Microsoft Remote Desktop Protocol be used Al6a For setup debugging and configuration for these servers it is recommended to use TeamViewer This software allows a remote client to connect any one desktop window at a time In comparison VNC in comparison will also capture and stream with the NVIDIA GPU accelerated driver but for baremetal systems it will capture all desktop windows This adds additional overhead and results in reduced performance when streaming The overhead for capturing one desktop window in TeamViewer is significantly less vs capturing all windows desktops with the VNC solution GRID SDK FAQ PG 06183 001 6 Note TeamViewer uses a proprietary compression format for remote streaming The NVENC H 264 hardware engine is not used by TeamViewer For streaming applications and games not using H 264 for compression TeamViewer is the recommended solution for streaming the desktop For the best experience on GRID use NVENC with H 264 compression for the best remote streaming experience A16b Can I use Microsoft Windows Remote Desktop While it is more efficient in terms of bitrate and performance of remote graphics in comparison to VNC Microsoft Remote Desktop uses a proprietary software based graphics driver that does not support al
17. wly created DirectX device GRID SDK FAQ PG 06183 001 7 Q18 How do I enable NvFBC A18 Here are the NvFBC settings a Windows NvFBC needs to be enabled after a clean installation Installation Directory bin gt NvFBCEnable exe enable Use the NvFBCH264 SDK sample to capture the desktop to a H 264 file b Linux For the NVFBC GRID API functions please refer to the flag NVFBC_ CREATE CAPTURE SESSION PARAMS bWithCursor Q19 What are the optimal encoder settings for desktop applications streaming A19 With the preset setting LOW LATENCY HP a single GPU can encode 6 8 streams exact number depends on other settings such as RC mode LOW _LATENCY_HQ preset will give 4 6 streams exact number depends on other settings such as RC mode The HQ preset will give you slightly better quality but the performance will be lower than that of HP To get the ideal performance for each preset you can use PerfMSNEC sample application in the SDK GRID SDK exposes 3 video encoder presets LOW_LATENCY_HQ LOW LATENCY _HP and LOW LATENCY DEFAULT As the names suggest these presets are meant for encoding for low latency applications such as remote streaming cloud gaming etc Please refer to the API reference for more details on each parameter Q20 What are the recommended hypervisors if I am using virtualized environment A20 We recommend XenServer 6 2 with SP1 for a NMOS for stability a
Download Pdf Manuals
Related Search