Nvidia GTX 980 / GTX 970 (Compute performance added)

[Gig11gs]

OK we all know how GTX 750 preform and how efficient it is, and now the big boys are here based on the same architecture "Maxwell"

you will find all the info here

Compute performance











any one excited?

[Kgx Pnqvhm]

I've been counting the days.....

[atom]

Based on the specs posted on videocardz the number of cuda cores is a lot less than expected. Therefore, the raw MD5 cracking speed should at ~10.9 BH/s. As comparison, the reference 290x is doing ~11.7 BH/s. For SHA1, it will be around 3852 MH/s and the 290x is at 3720 MH/s.

Do not forget it's mostly the chipset instructions that gives the 980 it's comeback into crypto not the raw hardware power. Therefore, algorithms that can not make use of the instructions will be slower on the 980. For example DEScrypt, it will be around 100 MH/s and 290x is at 130 MH/s.

The TDP is much better, so it's 160W on the 980 and 290W on the 290x and where's less power consumed there's less heat produced.

My (theoretical) conclusion is that the speed is more or less equal to the speed of the 290x. But for sure the NV drivers will be better (and released more often). The TDP is better, the heat should be less. But most important for the hashcracking scene, the multihash performace forthe unsalted algorithms is ways better.

[ocl]

All these benchmark/review sites should add Hashcat to their list.

165watt TDP is indeed very good. Still waiting for the hc benchmarks.

[epixoip]

Looks like speed will be about equal to an R9 290, slightly slower than an R9 290X. But that TDP makes up for the difference in performance and then some. Better drivers and no x11 are icing on the cake.

[Wushu]

The fact that there is finally an nVidia card that will give the AMDs a run for their money is alone a major breakthrough. No longer will we have to choose between pure speed and actual working drivers!

I'm actually the most stoked about the power draw. The reduced heat will be a nice touch.

[Wushu]

What kind of calculations do you use to estimate your MD5 speed? Are you calculating instructions per second possible and then using the average number of instructions an MD5 requires to be calculated (I've seen that at 398 instructions)?

[epixoip]

We know the real-performance of the 750Ti, and since it is from the same GPU family, the 980 will scale linearly against it. Estimating the speed is simple arithmetic:

750Ti: 640 cores @ 1200 Mhz.
980: 2048 cores @ 1216 Mhz.

(2048 * 1216) / (640 * 1200) =~ 3.24

So the GTX 980 will be ~ 3.24x faster than the GTX 750Ti. Using the benchmarks from the 750Ti we can now determine what the speed will be like for each algorithm on the 980 (note: based on outdated benchmark figures!)

750Ti MD5 = 3022.7 MH/s * 3.24 = GTX 980 9793.5 MH/s
750Ti NTLM = 5185.9 MH/s * 3.24 = GTX 980 16802.3 MH/s
750Ti SHA1 = 1065.3 MH/s * 3.24 = GTX 980 3451.6 MH/s

These are not the latest numbers -- these are based on atom's benchmarks from May, which are surely outdated. But it illustrates how we can accurately estimate the speed for different GPUs (and CPUs) from the same family.

We can also do it by the instruction count as well, but I'm not sure what the instruction counts are at with LOP3.LUT. But just as an example, here's how we can e.g. estimate NTLM performance on the R9 290X without having one:

2816 cores * 1000000000 Hz / 130 instructions = 21.6 GH/s

Which is spot-on with the actual benchmark:

Speed.GPU.#2.: 21663.7 MH/s

[Kgx Pnqvhm]

Does the 4 GB of GDDR5 video buffer memory help in any way?

[epixoip]

it helps in that you can load large hash lists, and get more parallelism out of obese algorithms like scrypt. but perhaps more important than the quantity of ram is the fact that nvidia has a vastly superior memory controller than amd, so you won't see near as much of a performance hit on large lists like you do on amd.

this card is a major win no matter how you look at it.