Benchmarks from NVidia's GTX 980
#1
Finally, the GTX 980 arrived. Some really nice changes come with it.

More details here:
Let the numbers talk:

[Image: 290vs980.png]

I still have to do some more research on the new sm_52 chipset features. There is a good chance that there is some hidden potential we can use to gain out a bit more speed. That's all I can say for now.

Conclusion:

- The 290x is still in the lead from raw numbers
- The 980 is in the lead from from Speed:Watt ratio (TDP for 290x is 290W, TDP for 980 is 160-180W)

--
atom
#2
Hoooo, hope v1.31 can solve that error--"cuModuleLoad<> 301".

BTW, GTX980's Power loss seems really good....
#3
Which specific card did you use?
Some makers have factory overclocked versions, which might be good to know for comparison purposes.

And which ForceWare driver?
#4
(09-24-2014, 07:57 PM)Kgx Pnqvhm Wrote: Which specific card did you use?
Some makers have factory overclocked versions, which might be good to know for comparison purposes.

And which ForceWare driver?

Well it's a GAINWARD card, but it doesn't matter, it's a reference card with default clock speeds. The driver I used is ForceWare 343.22 which is the latest stable driver for Linux.
#5
(09-24-2014, 06:22 PM)underwater Wrote: Hoooo, hope v1.31 can solve that error--"cuModuleLoad<> 301".

BTW, GTX980's Power loss seems really good....

oclHashcat v1.30 does not support the gtx980 (and 970?). I had to upgrade to the latest cuda SDK to get the compiler to work.

No worries with oclHashcat v1.31 they will be supported.
#6
OK, some nice update. The benchmark above uses brute-force to meassure the performance. When I played a bit more with the card I did a benchmark using combinator-attack, the -a1 kernels. This kernel is also used with hybrid attack -a6 and -a7.

The picture is a bit different:

290x:
Speed.GPU.#1...: 5866.1 MH/s

980:
Speed.GPU.#1...: 5975.6 MH/s

This is using MD5. I hope there are more surprises like this waiting to be found Smile
#7
i enjoy seeing a nvidia card performing better than amd! Better performance on combinator is something Big Grin Does this also work for rule-based attacks? And holds this true for other algos, too, or just md5?
#8
(09-24-2014, 09:50 PM)atom Wrote:
(09-24-2014, 06:22 PM)underwater Wrote: Hoooo, hope v1.31 can solve that error--"cuModuleLoad<> 301".

BTW, GTX980's Power loss seems really good....

oclHashcat v1.30 does not support the gtx980 (and 970?). I had to upgrade to the latest cuda SDK to get the compiler to work.

No worries with oclHashcat v1.31 they will be supported.

Yes, I have tried v1.30 on gtx980 SLI, but it didn't work only show me a error Sad

Atom told me that v1.31 can solve this. That's happy to hear that Smile
#9
interested in 970 benchmarks. there's supposed to be a 20nm version coming later? wonder if it consumes even less watts. I might get a 20nm 970.
#10
New optimization found in -a1 mode:

980 old -a1:
Speed.GPU.#1...: 5975.6 MH/s

980 new -a1:
Speed.GPU.#1...: 6507.2 MH/s