GPU

[kukuruza]

Hi, @Lotuslisa, today is the deadline for the nvidia grant as far as I understand?

Also, shall we put together the criteria to buy a GPU?

1) @Lotuslisa, you said that one is not enough. I kind of agree because one will probably be always busy collecting data for training, non-stop. But they are expensive 2) you said you will ask your friends about which GPU to get. What did they say? Some time ago we looked together at the minimum Faster-RCNN requirements. The first in the list -- Titan with 12 GB memory starts at around $1000 3) you and @LynnaGui said there are requirements for the computer itself as well. So what are they?

[kukuruza]

The highlights from the quote from Utilities and our updates:

• Dell T630 Server Chassis
• similar to Xeon E5 1650 6 core 3.5 GHz
• RAM 32 GB
• combination of 2 TB HDD and 128 GB SSD

Please take the chance to make your input, if any.

[Lotuslisa]

GPU and Computer System Purchase

1) Do not use Quadro series GPU. Because Quadro GPU is mainly for unlinear computation. 2) For deep convolutional neural networks, it’s better to choose Telsa series GPU. For the PS, it’s better to choose GTX series. 3) K40 is better for double precision computing. Titan x is faster than k40, but with single precision computing. For car detection we can just use single precision computing. 4) The power of GPU is high. so we should have the power supply more than 500w. 5) The chipsets board must support PCIe3.0 6) the size of RAM influence the performance a lot. it should be more than 32G

[kukuruza]

3) According to Faster-RCNN benchmarks, indeed Titan X is faster for deployment. Titan X is also the only card from GTX series that has enough memory for the VGG model. 4) Right. I didnt mention that, ECE IT support suggested 1100 Wt power suply 6) Ok, so let it be 32 GB then.

[kukuruza]

Closing, since the server works.