For GPU computing needs I like systems with lots of PCIe slots.
Here’s how I gauge the CUDA performance of my GPUs per dollar spent, from lowest cost per TE to highest:
1) EVGA GTX 480 SC /1.5Gig (G) = TE of .613; Price = ~$200 (on Ebay) [ for each $1000 spent or with
5 of them you’d get a TE of 3.065 { 1000 / 200 = 5; 5 * 0.613 = 3.065 }];
2) EVGA GTX 590 Classified (C) / 1.5G per GPU = TE of 1.13; Price = ~ $400 (on Ebay) [ for each $1000 spent or with
2.5 of them you’d get a TE of 2.825 { 1000 / 400 = 2.5; 2.5 * 1.13 = 2.825 }];
3) EVGA GTX 580C / 3G = TE of .594; Price = ~ $300 (on Ebay) [ for each $1000 spent or with
3.33 of them you’d get a TE of 1.807 { 1000 / 300 = 3.333; 3.333 * 0.594 = 1.980 }];
4) EVGA GTX 780 Ti Superclock (SC) ACX / 3G = TE of 1.319; Price = $730 from EVGA [ for each $1000 spent or with
1.37 of them you’d get a TE of 1.807 { 1000 / 730 = 1.37; 1.37 * 1.319 = 1.807 }];
5) Galaxy 680 / 4G = TE of .593; Price = ~ $460 for comparable GPU at NewEgg [ for each $1000 spent or with
2.174 of them you’d get a TE of 1.289 { 1000 / 460 = 2.174; 2.174 * 0.593 = 1.289 }];
6) EVGA GTX 690 / 2G per GPU = TE of 1.202; Price = $1,000 from EVGA [ for each $1000 spent or with
1.0 of them you’d get a TE of 1.202 { 1000 / 1000 = 1.0; 1.0 * 1.202 = 1.202 }];
7) EVGA GTX Titan SC / 6G = TE of 1.185; Price = $1,020 from EVGA [ 1.185 * 1000 = $1185]; and
8) Titan Reference Design that Bare Feats tested = TE of 1.0; Price= $1,000 from EVGA [ for each $1000 spent or with
1.0 of them you’d get a TE of 1.0 { 1000 / 1000 = 1.0; 1.0 * 1.0 = 1.0 }].
Obviously, other factors will come into play. These can include your budget, the number of available double wide PCIe slots that you have, as well as your needs, which can be influenced by business demands such as typical project size and your work style such as number and size of textures. Your clientele may have a bearing on which card(s) appear(s) to be the best value for your money.
Additionally, this February we could see the introduction of the GeForce GTX 790 that may feature 10GB of memory, with dual 320b interfaces and 2496 CUDA cores per GPU [and cards like the GTX 860 - based on the mid-range GM107 GPU] [
http://videocardz.com/48610/nvidia-maxwell-details-revealed-ces-2014 ].
Lastly, there will soon be more credible portable GPU computing options. According to Videocardz, in February, NVIDIA will be releasing their next mobile flagship - GeForce GTX 880M, using the GK104, with 1536 CUDA cores, with 8GB of memory [
http://videocardz.com/48633/nvidia-geforce-gtx-880m-rebranded-gtx-780m-8gb-memory ]. For those more interested in better portable OpenCL compute performance, AMD officially introduced at CES 2014 its R7 M200 series, though currently such mobile graphics cards are only available in certain systems, including Alienware, Clevo, Lenovo and MSI [
http://videocardz.com/48667/amd-radeon-r9r7r5-m200-series-officially-announced ].
