Sounds good but cannot find a vendor with these in stock, seems like a good value.
Any benchmarks out there?
They seem to be 3~8% slower (depending on the parameters) than a TITAN X in my compute-bound application, and 6GB memory is ample for most tasks. Considering they’re 35% cheaper than TITAN Xes, I’d say they’re very competitive FP32 CUDA cards.
Thanks for the info.
Are the clock speeds of the two the same?
For example I think the GTX Titan Black generally ran at a lower clock speed than the GTX 780ti.
This will also depend on the manufacturer of the that particular GPU model, as a company like EVGA will add an additional ACX cooler then overclock about 10%
Also what about memory bound applications, any noticeable difference?
Tricky question! I actually bought three Gigabyte GTX 980 Tis (reference cards) and a Gigabyte GTX Titan X (also reference card). Their actual core clock all differ under full load (no OC, default voltage), presumably due to different ASIC quality, despite the typical boost clock all being the same at 1076 MHz. For the 980 Tis, higher ASIC quality ones have higher actual clock, but they all range in between 1189 to 1215 Mhz. The Titan X runs at 1177 Mhz at full load. So no, they do not run at same clock speed at default. The peak GFLOPS at default clocks are: Titan X @ 6866, 980 Tis @ 6478 to 6569 (again, higher ASIC quality ones have higher GFLOPS).
I also did some slight OC to force them all run at 1220 Mhz, the peak GFLOPS were: Titan X @ 7073, 980 Tis @ 6561 to 6583.
All of my 980 Tis can be overclocked +300 Mhz @ 1501 to 1527 Mhz at default voltage, raising the GFLOPs to 8077 to 8203, compared against the Titan X @ 1477 Mhz giving 8609 GFLOPS.
I haven’t tested any memory bound applications as I don’t have any on hand, but I would guess there wouldn’t be much difference as they are of same architecture and the advertised bandwidth are the same. Are there any free CUDA memory bandwidth benchmark software I can use?
Another thing I’ve noticed is that algorithms that I’ve optimized for Titan X need re-tuning of parameters to run at maximum speed on 980 Ti and vice-versa. This stroke me as odd as I was under the impression they’re very similar GPUs. On some parameter sets I even see the Titan X performing worse than the 980 Ti. I first thought it was due to SM workload balancing issues as the Titan X has 24 SMs and the 980 Ti has 22, but that wasn’t the case. Profiling shows the biggest difference seems to be warp issue efficiency on my program across different cards. I’m still nailing down the exact reason.
How about the cuda sample code bandwidthTest ?
The device-to-device copy reported number should be a reasonable proxy for relative comparison of different GPUs.
(some GPUs may need to have “boost clocks enabled” for comparison purposes. I don’t think that’s the case with these GPUs.)
They all clock @ 7010 Mhz, and the D to D transfer rates are around (±0.2%) 249,500 MB/s for all four of my cards. I think it’s safe to say the memory performance of Titan X and the 980 Ti are the same, provided Nvidia doesn’t pull another 970.
Interestingly CUDA-Z seems to show very different DtoD transfer rates, I’ll take it as the result of different implementation.
tt
Thank you for the detailed information.
I have an application which needs slightly higher performance than what I was achieving on a not-overclocked reference GTX Titan X.
Trying to figure if I should get a factory superclocked Titan X or the GTX 980ti.
As far as a good memory bandwidth test, an engineer who frequents this board, Jimmy Pettersson, submitted some an excellent CUDA memory bandwidth test which I use to gauge “best-case-scenario” bandwidth.
Here is a link to the source code, which(if you are interested) you can just copy-paste-compile run;
My output on my Titan X was this;
GeForce GTX TITAN X @ 336.480 GB/s
N [GB/s] [perc] [usec] test
1048576 156.60 46.54 26.8 Pass
2097152 198.52 59.00 42.3 Pass
4194304 227.90 67.73 73.6 Pass
8388608 255.37 75.90 131.4 Pass
16777216 269.79 80.18 248.7 Pass
33554432 277.85 82.58 483.1 Pass
67108864 282.06 83.83 951.7 Pass
134217728 284.33 84.50 1888.2 Pass
Non-base 2 tests!
N [GB/s] [perc] [usec] test
14680102 269.23 80.01 218.1 Pass
14680119 269.36 80.05 218.0 Pass
18875600 266.36 79.16 283.5 Pass
7434886 155.41 46.19 191.4 Pass
13324075 240.67 71.53 221.4 Pass
15764213 252.76 75.12 249.5 Pass
1850154 60.04 17.84 123.3 Pass
4991241 139.61 41.49 143.0 Pass
Press any key to continue . . .
The champ in this test is still the GTX 780ti which was about 3% higher than the above numbers. The code was geared towards the Kepler, so maybe there are Maxwell optimizations.
Here are sample results on my four cards (factory clock, voltage). Judging from what I see across several runs on the same card, I would say the memory bandwidth differences between different cards below are mostly fluctuation.
GeForce GTX TITAN X @ 336.480 GB/s
N [GB/s] [perc] [usec] test
1048576 148.24 44.06 28.3 Pass
2097152 192.36 57.17 43.6 Pass
4194304 221.20 65.74 75.8 Pass
8388608 252.56 75.06 132.9 Pass
16777216 267.58 79.52 250.8 Pass
33554432 275.92 82.00 486.4 Pass
67108864 280.58 83.39 956.7 Pass
134217728 282.34 83.91 1901.5 Pass
Non-base 2 tests!
N [GB/s] [perc] [usec] test
14680102 268.26 79.72 218.9 Pass
14680119 268.79 79.88 218.5 Pass
18875600 265.64 78.95 284.2 Pass
7434886 156.25 46.44 190.3 Pass
13324075 240.50 71.48 221.6 Pass
15764213 252.23 74.96 250.0 Pass
1850154 60.86 18.09 121.6 Pass
4991241 141.14 41.95 141.5 Pass
Press any key to continue . . .
GeForce GTX 980 Ti @ 336.480 GB/s
N [GB/s] [perc] [usec] test
1048576 135.25 40.19 31.0 Pass
2097152 183.76 54.61 45.6 Pass
4194304 220.88 65.64 76.0 Pass
8388608 247.68 73.61 135.5 Pass
16777216 263.54 78.32 254.6 Pass
33554432 273.81 81.38 490.2 Pass
67108864 280.07 83.23 958.5 Pass
134217728 282.91 84.08 1897.7 Pass
Non-base 2 tests!
N [GB/s] [perc] [usec] test
14680102 267.88 79.61 219.2 Pass
14680119 268.09 79.68 219.0 Pass
18875600 265.39 78.87 284.5 Pass
7434886 156.28 46.45 190.3 Pass
13324075 239.71 71.24 222.3 Pass
15764213 252.25 74.97 250.0 Pass
1850154 60.76 18.06 121.8 Pass
4991241 140.38 41.72 142.2 Pass
Press any key to continue . . .
GeForce GTX 980 Ti @ 336.480 GB/s
N [GB/s] [perc] [usec] test
1048576 123.11 36.59 34.1 Pass
2097152 170.48 50.67 49.2 Pass
4194304 211.95 62.99 79.2 Pass
8388608 241.84 71.87 138.7 Pass
16777216 259.46 77.11 258.6 Pass
33554432 272.92 81.11 491.8 Pass
67108864 279.12 82.95 961.7 Pass
134217728 282.44 83.94 1900.8 Pass
Non-base 2 tests!
N [GB/s] [perc] [usec] test
14680102 265.83 79.00 220.9 Pass
14680119 265.75 78.98 221.0 Pass
18875600 263.62 78.35 286.4 Pass
7434886 156.29 46.45 190.3 Pass
13324075 237.27 70.52 224.6 Pass
15764213 250.34 74.40 251.9 Pass
1850154 60.20 17.89 122.9 Pass
4991241 139.18 41.36 143.4 Pass
Press any key to continue . . .
GeForce GTX 980 Ti @ 336.480 GB/s
N [GB/s] [perc] [usec] test
1048576 138.05 41.03 30.4 Pass
2097152 182.31 54.18 46.0 Pass
4194304 221.25 65.75 75.8 Pass
8388608 248.17 73.75 135.2 Pass
16777216 261.92 77.84 256.2 Pass
33554432 273.67 81.33 490.4 Pass
67108864 280.10 83.24 958.4 Pass
134217728 282.96 84.09 1897.3 Pass
Non-base 2 tests!
N [GB/s] [perc] [usec] test
14680102 266.57 79.22 220.3 Pass
14680119 268.35 79.75 218.8 Pass
18875600 265.53 78.91 284.4 Pass
7434886 158.88 47.22 187.2 Pass
13324075 240.36 71.43 221.7 Pass
15764213 252.70 75.10 249.5 Pass
1850154 61.33 18.23 120.7 Pass
4991241 141.36 42.01 141.2 Pass
Press any key to continue . . .