The NVidia Graphics Card Specification Chart contains the specifications most used when selecting a video card for video editing software and video effects software. Chart by David Knarr.
Contents.Field explanations The fields in the table listed below describe the following:. Model – The marketing name for the processor, assigned by Nvidia. Launch – Date of release for the processor. Code name – The internal engineering codename for the processor (typically designated by an NVXY name and later GXY where X is the series number and Y is the schedule of the project for that generation). – Fabrication process. Further information:.
All models are made via 220 nm fabrication process. All models support 7.0 and 1.2.
All models support hardware Transform and Lighting (T&L) and Cube Environment MappingModelLaunchCore clock Memory clock Core config 1MemoryMOperations/sMPixels/sMTexels/sMVertices/sSize Bandwidth (/s)Bus typeBus width GeForce 256 SDROctober 11, 1999NV10220 nmAGP 4×PCI1201664:4:0322.656SDR128GeForce 256 DDRFebruary 1, 44.8DDR. 1::GeForce2 series. Further information:.
All models are made via 150 nm fabrication process. All models support 8.0 and 1.3. All models support 3D Textures, Lightspeed Memory Architecture (LMA), nFiniteFX Engine, Shadow BuffersModelLaunchCore clock Memory clock Core config 1MemoryMOperations/sMPixels/sMTexels/sMVertices/sSize Bandwidth (/s)Bus typeBus width GeForce3 Ti200October 1, 2001NV20150 nmAGP 4×PCI1752004:1:8:06.4DDR128GeForce3February 27, 50647.36GeForce3 Ti500October 1, 60641288. 1:::GeForce4 series. Further information: and.
1::. 2 To calculate the processing power see.ModelLaunchTransistors (million)Die size (mm 2)Core config 1Clock rateMemory configurationSupported versionProcessing power 2(Watts)CommentsCore Shader Memory Pixel (/s)Texture (/s)Size Bandwidth (/s)DRAM typeBus width GeForce G 100March 10, 2009G9865 nm21086PCIe 2.0 ×168:8:02.154.35128.0DDR26410.03.322.435OEM productsGeForce GT 120G96bTSMC 55 nm3141218004.48.816.012889.650GeForce GT 196?
3675GeForce GT 1400.420.8512 102457.6GDDRGeForce GTS 16011.141GeForce 200 series. Further information: andMemory bandwidths stated in the following table refer to Nvidia reference designs.
Further information: and. 1::. 2 To calculate the processing power see. 3 Each SM in the GF110 contains 4 texture filtering units for every texture address unit.
The complete GF110 die contains 64 texture address units and 256 texture filtering units. Each SM in the GF114/116/118 architecture contains 8 texture filtering units for every texture address unit but has doubled both addressing and filtering units. 4 Internally referred to as GF104B. 5 Internally referred to as GF100B.
6 Similar to previous generation, GTX 580 and most likely future GTX 570, while reflecting its improvement over GF100, still have lower rated TBP and higher power consumption, e.g. GTX580 (243W TBP) is slightly less power hungry than GTX 480 (250W TBP).
This is managed by clock throttling through drivers when a dedicated power hungry application is identified that could breach card TBP. Further information: andThe GeForce 700 series for desktop. The GM107-chips are -based, the GKxxx-chips. 1::. 2 Max Boost depends on ASIC quality.
For example, some GTX TITAN with over 80% ASIC quality can hit 1019 MHz by default, lower ASIC quality will be 1006 MHz or 993 MHz. 3 Kepler supports some optional 11.1 features on 110 through the Direct3D 11.1 API, however Nvidia did not enable four non-gaming features to qualify Kepler for level 111.
4 The GeForce GT 705 (OEM) is a rebranded GeForce GT 610, which itself is a rebranded GeForce GT 520. 5 The GeForce GT 730 (DDR3, 64-bit) is a rebranded GeForce GT 630 (Rev. 2).
6 The GeForce GT 730 (DDR3, 128-bit) is a rebranded GeForce GT 430 (128-bit). 7 The GeForce GT 740 (OEM) is a rebranded GeForce GTX 650. 8 The GeForce GTX 760 Ti (OEM) is a rebranded GeForce GTX 670. 9 To calculate the processing power see, or. 10 As a Kepler GPC is able to rasterize 8 pixels per clock, fully enabled GK110 GPUs (780 Ti/TITAN Black) can only output 40 pixels per clock (5 GPCs), despite 48 ROPs and all SMX units being physically present. Main:: (Streaming Multiprocessors). ^ Base clock, Boost clock.
To calculate the processing power see. Pixel fillrate is calculated as the number of ROPs multiplied by the respective core clock speed. Texture fillrate is calculated as the number of TMUs multiplied by the respective core clock speed. ^ For accessing its memory, the GTX 970 stripes data across 7 of its 8 32-bit physical memory lanes, at 196 GB/s. The last 1/8 of its memory (0.5 GiB on a 4 GiB card) is accessed on a non-interleaved solitary 32-bit connection at 28 GB/s, one seventh the speed of the rest of the memory space.
Because this smaller memory pool uses the same connection as the 7th lane to the larger main pool, it contends with accesses to the larger block reducing the effective memory bandwidth not adding to it as an independent connection could.GeForce 10 series. Further information: and. Supported: 12 (121), 4.6, 1.2, 1.1 and 7.5.
Unlike previous generations the RTX Non-Super (RTX 2070, RTX 2080, RTX 2080 Ti) Founders Edition cards no longer have reference clocks, but are 'Factory-OC'. Further information:The GeForce 600M series for notebooks architecture. The processing power is obtained by multiplying shader clock speed, the number of cores, and how many instructions the cores can perform per cycle. 1::ModelLaunchFab Core config 1Clock speedMemorySupported versionProcessing power 2(Watts)NotesCore Shader Memory Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width GeForce 610MDecember 2011GF119 (N13M-GE)40PCIe 2.0 ×1648:8:04204814.4DDR3.0812OEM. Rebadged GT 520MXGeForce GT 620MApril 2012GF117 (N13M-GS)2812504.44015OEM.
Die-Shrink GF108GeForce GT 625MOctober 201214.464GeForce GT 630MApril 2012GF108 (N13P-GL)GF.63.210.712.828.832.0DDR3GDDR5128.233GF108: OEM. Rebadged GT 540MGF117: OEM Die-Shrink GF108GeForce GT 635MGF106 (N12E-GE2)GF1164016.23628.843.2DDR.2388.835GF106: OEM. Further information:The GeForce 700M series for notebooks architecture. The processing power is obtained by multiplying shader clock speed, the number of cores, and how many instructions the cores can perform per cycle. 1::ModelLaunchFab Core config 1Clock speedMemorySupported versionProcessing power 2(Watts)NotesCore Shader Memory Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width GeForce 710MJanuary 2013GF11728PCIe 2.0 ×16160018003.24814.4DDR3.212OEM.
About 115% of Mobile 620 & Desktop 530 GeForce GT 720MApril 1, 620003.8.0360.19?OEM. Further information:The GeForce 800M series for notebooks architecture. Further information:The GeForce 900M series for notebooks architecture.
Further information:. 1:::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported version(Watts)NotesPixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 540NV43GL90PCIe ×04:8:8:8.8GDDR31289.0c2.135Quadro FX 1400NV06005:12:12:82.84.219.2DDR25675Stereo display,Quadro FX 3400NV45GL/ NV409006:16:16:128.8GDDR3101Quadro FX 3450NV510005:12:12:125.15.132.083Quadro FX 4400NV45GL A3/ NV010506:16:16:133.7110Quadro FX (x500) series. Further information:. 1:::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported version(Watts)NotesPixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 350G72GL90PCIe ×03:4:4:6.48DDR2649.0c2.121Quadro FX 550NV43GL:8:8:82.882.8812.8GDDR312825Quadro FX 1557.565Quadro FX 3500G71GL7:20:20:167.529.442.280Stereo display,Quadro FX 45508.28Quadro FX 4500X22× 8.0445Stereo display, GenlockQuadro FX 4500 SDI8:24:24:167.526Quadro FX 5570011.2Stereo display, GenlockQuadro FX 5500 SDI104Quadro FX (x600) series. Further information:. 1:::. 2::ModelFab Core clock Shader clock Memory clock Core config 12MemoryProcessing power Supported version(Watts)NotesPixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 560G73GL90PCIe ×005:12:12:19.2GDDR31289.0c2.130Quadro FX 4600 2G40024345-Stereo display, GenlockQuadro FX 4600 SDI 25154Quadro FX 5600 2PCIe 2.0 ×6001.4.8518.4171Quadro FX (x700) series.
Further information:. 1::. 4 Each SM in the Fermi architecture contains 4 texture filtering units for every texture address unit. Further information:.
1::: streaming multiprocessorsModelFabCore clockShader clockMemory clockCore config 1MemoryProcessing power Supported version(Watts)NotesPixel(/s)Texture(/s)SizeBandwidth(/s)Bus typeBus widthQuadro M2000GM20628PCIe 3.0 ×1653(6612)768:48:32:637.856.64105.8GDDR51281812.556.612.14.61.11.25.275Four DisplayPort 1.2aQuadro M4000GM503(6012)1664:1.283.2662.483.2120Quadro M5)2048:1.2134.4211.4150Four DisplayPort 1.2a, One DVI-IQuadro M6000GM653(6612)3072:1.810250Quadro Pxxx series. ^ To calculate the processing power see,. A number range specifies the minimum and maximum processing power at, respectively, the base clock and maximum boost clock.
Core architecture version according to the programming guide. ^ GPU Boost is a default feature that increases the core clock rate while remaining under the card's predetermined power budget.
Multiple boost clocks are available, but this table lists the highest clock supported by each card. ^ Specifications not specified by Nvidia assumed to be based on the GTX.
^ Specifications not specified by Nvidia assumed to be based on the. ^ Specifications not specified by Nvidia assumed to be based on the Quadro FX 5800. ^ With ECC on, a portion of the dedicated memory is used for ECC bits, so the available user memory is reduced by 12.5%. 4 GB total memory yields 3.5 GB of user available memory.)Mobile Workstation GPUs Quadro Go (GL) & Quadro FX Go series.
Further information:Early mobile Quadro chips based on the Geforce2 Go up to Geforce Go 6. Further information:GeForce 7-Series based. 1:::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported version(Watts)Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 350MG72GLM90PCIe 1.0 ×03:4:4:14.4GDDR31289.0c2.115Quadro FX 1500MG71GLM8:51232256Quadro FX 2520081238.4Quadro FX 35.213.8Quadro FX (x600M) series. Further information:GeForce 8-Series (except FX 560M and FX 3600M) based. First Quadro Mobile line to support DirectX 10.
1::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported version(Watts)NotesPixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 360MG86M80PCIe 1.0 ×0016:8:9.6DDR26438.410.03.317Based on the GeForce 8400M GSQuadro FX 560MG73GLM905005005:1219.2GDDR31289.0c2.135?7600GS based?Quadro FX 1600MG50160025.612010.03.350?Quadro FX 3600MG:32:16124036070Based on the GeForce 8800M GTXQuadro FX (x700M) series. Further information:. 1::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported version(Watts)Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 370MG98M65PCIe 1.0 ×2008:4:9.6GDDR36433.610.03.320Quadro FX 570MG0140032:16:22.412891.245Quadro FX 770MG.6119.035Quadro FX 17510148.850Quadro FX 2700MG48:24:168.4812.72.865Quadro FX 3700MG:64:168.8875Quadro FX (x800M) series. Further information:The last DirectX 10 based Quadro mobile cards. 1::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported version(Watts)Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro FX 380MGT218M40PCIe 2.0 ×60016:8.8GDDR.13.325Quadro FX 880MGT248:16:4.2435Quadro FX 1800MGT2:24:83.610.825.635.2GDDR3GDDR5233.2845Quadro FX 2800MG502000664GDDR325636010.075Quadro FX 381.843.2648.192100Quadro (xxxxM) series. Further information:. 1::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported versiontechnology(Watts)NotesPixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro K500MGK10728PCIe 3.0 ×00192:16:86.8.8DDR364326.411.04.5Yes35Quadro K10:16:16.8128326.445Dell Precision M4700Quadro K20.9223.Dell Precision M4700Quadro K3000MGK8005.9331.3989.6GDDR52Dell Precision M6700Quadro K40.2100Dell Precision M6700Quadro K500001344:112:3222.597.73Dell Precision M6700Quadro (Kx100M) series.
Further information:. 1::ModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported versiontechnology(Watts)Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro M500MGM10828PCIe 3.0 ×1)384:24:8(3 SMM)204840.1GDDR56412.14.5Yes25Quadro M600MGM107384:24:16(3 SMM)80.212830Quadro M1000M512:32:16(4 SMM)40Quadro M208640:40:16(5 SMM)Quadro M3000MGM2041024:64:32(8 SMM)160.425675Quadro M4000M1280:80:48(10 SMM)100Quadro M50536:96:64(12 SMM)61.655.3Quadro (Mx200) series. Further information:.
1::: streaming multiprocessorsModelFab Core clock Shader clock Memory clock Core config 1MemoryProcessing power Supported versiontechnology(Watts)Pixel (/s)Texture (/s)Size Bandwidth (/s)Bus typeBus width Quadro P500GP10814PCIe 3.0 ×16256:840GDDR56475012.14.5Yes18Quadro P600GP107384:25Quadro P1000GP107512:696160040Quadro P2000GP107768:0240050Quadro P3000GP106161280:Quadro P4000GP1041792:10Quadro P4000Max-Q1792:1Quadro P50002048:1384100Quadro (Px200) series.
I struggled at first to get Tensorflow installed and working correctly for the NVIDIA GPUs. After many trial and errors, I found a consistent way to get it to work. I am sharing the step-by-step guide to getting Tensorflow working on your CentOS 7 distribution, using the NVIDIA GPUs.My environments are multiple workstations that use various combinations of Nvidia Titan X, Xp, RTX, and Tesla GPUs. Typically, they are installed in the multiple of 2s. Two to four GPU workstations are most common where I support the systems.While this article deals specifically with version 10.0 of CUDA used for Tensorflow r1.13, the general steps apply for other version.Make sure to consult the CUDA compatibility table found on NVIDIA site.Generally speaking, your CUDA toolkit version MUST match the NVIDIA Video Graphics Driver version. It is very easy to mix up the wrong versions. I have done it myself and spent hours and days fixing it.
In order to use TensorFlow on your workstation, there are a few assumptions and requirements.Assumptions1.The workstation you are going to use TensorFlow has CentOS 7 or Red Hat Enterprise Linux Workstation 7; Ubuntu is often the most often used examples on the Internet, at Brown, we tend to use Red Hat and CentOS for most work, especially if you require support from the CIS or the division Systems Manager.2.The workstation has one or more of Nvidia CUDA compatible GPUs already installed in appropriate PCI-e slots in the chassis. For the list of CUDA compatible GPU devices, see3.Whenever possible, RPM or yum and PIP are the primary means used to install the required software and modules4.Python is your programming language for TensorFlow5.Anaconda installer, i.e.
Conda, is not going to be used for setting up the TensorFlow programming environment; you are welcome to use use Anaconda, however, please make sure that your installation of Anaconda does not interfere with the master installation of Python, PIP and TensorFlow.6.Version of TensorFlow that is going to be installed is tensorflow-gpu – the GPU enabled version of TensorFlow. There is a CPU version of TensorFlow, named tensorflow. You do not want to use this version if you are going to use GPU. Before installing TensorFlow on the workstation, several requirements must be satisfied.1.As noted above, you must have at least one CUDA compatible NVIDIA GPUs in the correct PCI-e 3.0 x16 slots.2.Install the correct, compatible version of the video graphics driver for the GPUs.
This is often done at the same time as the CUDA toolkit installation, however, it can be installed separately.3.Install the correct, compatible version of the CUDA toolkit (either 9.0 or 10.0)4.Install the correct, compatible version of Python (either 2.7 or 3.6 – 3.6 is recommended)5.Install the correct, compatible version of PIP that is designed to work with the version of PythonSee the detailed installation instruction later in this document. You can skip this step if you already have Python 3.6 and PIP 3.6 installed on your workstation.Whenever possible, use rpm and yum to perform the installation of packages and software modules. Yum will handle all the dependency checks and make your life much easier to manage the package installation and upgrades.You can, however, still use the vendor provided shell scripts or compile the modules from the source code. The shell scripts and the source code methods do not provide you with good dependency checking and can make your life difficult when TensorFlow refuses to work due to incompatible software and libraries. If you install CUDA and Tensorflow using the vendor provided scripts, then the RPM will not know what packages are actually installed. It may cause conflict, especially during the package update time.As root or use sudo to install Python 3.6 (not 3.7 – has been shown to cause compatibility issue for CUDA 9.) If you download the RPM directly from Python.org, make sure that it is version 3.6, not version 3.7.Use yum install install Python 3.6.
Note that there are version 3.4 and 3.6 available on the public repository. Make sure to specify the version 3.6 by using python36, not python34 when using yum.$ sudo yum install python36Since CentOS 7 uses Python 2.7 as the default Python version to manage dependent packages – such as yum, you want to leave it alone. You can have version 2.7 and 3.6 co-existing on the same workstation. They are typically installed in /bin directory, or sometimes in /usr/local/bin. However, to distinguish the two versions, it is recommended to create a symbolic link to the version 3.6 for convenience.Test which version is the default Python.$ sudo python -versionMost likely, the version displayed will be 2.7.x.
This is fine, since most CentOS tools will depend on having the default version of Python as 2.7.x.To use TensorFlow, you need to choose either 2.7 or 3.6 version of Python. The version 3.6 is preferred, as it is the future of Python.
Test which version of 3.6 you have in your $PATH.$ which python36Most likely, the answer will come back as /bin/python36. This is where the Python 3.6 is installed by default when you use yum to perform the package installation. Please make certain that the PATH environment variable is set so that your Python version is used.Create a symbolic link to Pyhon3.6 for convenience, eliminating a need to keep on typing python3.6.$ ln -s /usr/bin/python3.6 /usr/bin/python3Verify that the symlink is working correctly.$ python3 -versionThe above command should tell you which version of Python you are running.Now, let’s install PIP to go with Python 3.6. Once you met all the above requirements, then you are finally ready to install Tensorflow.
Please make sure you are installing the correct version of TensorFlow – “tensorflow” is a CPU only version.