Search the Community
Showing results for tags 'ubuntu mate 16.04'.
Found 1 result
NOTE: This post has been updated to reflect the latest state of this implementation... Hello mates, I am delighted to share a bit of my new successful implementation... After fighting my way thru previous EGPU implementations using several Linux distributions. From Ubuntu Mate 14 & 15 to Linux Mate and Centos 5 and 6. I only documented one of them. I had to share this experience, mostly because I am amazed by what the community behind Ubuntu Mate 16.04 has achieved. So bear with me. System Specs Lenovo T430 Intel Core i5-3320m at 2.6 Ghz 8 GB DDR3L 12800 16 GB DDR3L 12800 Intel HD 4000 EGPU: Zotac GeForce GTX 750 1GB EVGA GeForce GTX 950 SC+ 2GB KFA2 GeForce GTX 970 OC Silent "Infin8 Black Edition" 4GB EXP GDC v8.3 Beast Express Card Seasonic 350 watts 80+ bronze Display: Internal LCD 1600x900 Dell UltraSharp 2007FP - 20.1" LCD Monitor Procedure: I prepared the hardware as usual. Feeding power to the Beast adapter using the PSU. Plugging these into the laptop's ExpressCard slot. The installation of Ubuntu Mate 16.04 used is only a couple of weeks old and is loaded only with a full stack of Python and web tools I need. For the integrated graphic card, stock open source drivers are used. For the EGPU... I was ready to perform the usual steps, disable nouveau drivers, reboot switch to run level 3, install the cuda drivers, etc. But.. Following the advice read on a Ubuntu/Nvidia forum, and very sceptical, I installed the most recent proprietary drivers for my card. Reboot. Boom! I am done. Even functionality previously not available in Linux is now available... As you can see from the last screenshot the drivers now report what processes are being executed on the GPU, that was something reserved previously to high-end GPUs like Teslas. That screenshot also shows the evidence of the computation being performed in the GPU while the display is rendered in my laptop's LCD. This screenshot also shows how the proprietary driver can now display the GPU temperature as well as other useful data. For those of you into CUDA computing, I can report CUDA toolkit 7.5 is now available in the Ubuntu repository and also installs and performs without any issue. I went from zero to training TensorFlow models using the GPU in 30 minutes or so. Amazing! I could expand this post if anyone needs more info, but it was very easy. Cheers! After upgrading the GPU two times, my system is now capable of handling Doom fairly easy. Now some benchmark results. RAM eGPU PCIe gen 3d Mark 11 3dm11 Graphics 8 GB GTX 750 2 P3 996 4 095 16 GB GTX 750 2 P3 994 4 094 16 GB GTX 950 1 P5 214 7 076 16 GB GTX 950 2 P5 249 7 709 16 GB GTX 970 1 P7 575 11 202 16 GB GTA 970 2 P8 176 12 946 Now, the difference between Gen 1 and Gen 2 might not seem relevant from the results in the table. But playing Doom there is a difference of around 15 fps on average between both modes. This brief difference is even more noticeable during intense fights.