Nividia Titan

[John P. Myers]

This benchmark seems to be correct: http://www.setiusa.us/attachment.php...7&d=1360539612

As suspected, the 512bit interface seems to be incorrect.

Finally, we have some solid information about the upcoming flagship model from NVIDIA. Our sources confirm that the almighty Titan is set to be launched on February 18th in very limited quantities.

Forget about the 512-bit interface though. The GTX Titan will be based on GK110 GPU with 2688 CUDA cores. There will be 224 texture mapping and 48 raster operating units. The reference board will almost without a doubt be equipped with 6GB of GDDR5 memory across 384-bit interface.

First leaks suggested that the core will be clocked at 732 MHz. DonanimHaber has reported reliable, reliable as can be at this stage, information about the texture fill rate, which apparently comes in at 288 GT/s. That’s faster than the GTX 690′s 234 GT/s. Furthermore, the site is reporting that the GTX Titan would have computing power of 4.5 TFLOPS. If the provided numbers are correct we are in the range of 800-900 MHz core clock.

The card will definitely look better than the GTX 690, although the design will mirror it’s dual-gpu brother. The GeForce GTX Titan will be covered with the magnesium alloy, while the whole card will be metallic silver. There will only be a reference design, so no custom models. Additionally, in the first batch there may only be ASUS and EVGA cards available.

Expect reviews on the 19th and realistic availability around the 24th

[Crazybob]

So do you think I could get a couple to try out? Drooooolllll.................

[John P. Myers]

The company castrated the Double Precision, and you can expect great Single Precision performance (2,688 CUDA cores times 875 MHz should result in around 4.5 TFLOPS SP from a single chip). Double-precision follows the Kepler tradition of 1/24 Single Precision performance. Yes, 4.5 TFLOPS SP and 196 GFLOPS DP, nicely protecting Tesla K20/K20X and the upcoming Quadro K6000 products.

Welp, nevermind then. Keep buying AMD.

[DrPop]

Welp, nevermind then. Keep buying AMD.

What the heck? Why on earth does the marketplace let them get away with this crap? Crippled to the tune of 1/24th its true power??????????? Arrrrrgggghhhhh.

[zombie67]

Welp, nevermind then. Keep buying AMD.

Help me out. Can we get the equivalent numbers, or ratios, or whatever for the 580 and 680? I have no frame of reference here.

Also, at 1/24, how does that compare to (say) a 7970 at DP performance?

Finally, what projects require DP, or are significantly sped up with DP? Yes, MW does. GPUGRID uses almost no DP. Others?

[John P. Myers]

Help me out. Can we get the equivalent numbers, or ratios, or whatever for the 580 and 680? I have no frame of reference here.

Also, at 1/24, how does that compare to (say) a 7970 at DP performance?

Finally, what projects require DP, or are significantly sped up with DP? Yes, MW does. GPUGRID uses almost no DP. Others?

While you're right that DP isn't used for many projects, that was really the only selling point the Titan had for me - finally having an Nvidia GPU that didn't suck at double precision. Otherwise the $900 price tag for this thing is unfounded. Here's why:

Nvidia GTX Titan:
SP: 4.5 TFLOPS
DP: 196 GFLOPS
Price: $899
SP/$: $0.20 per SP GFLOPS
DP/$: $4.59 per DP GFLOPS

AMD HD 7970 GE:
SP: 4.096 TFLOPS
DP: 1024 GFLOPS
Price: $450
SP/$: $0.11 per SP GFLOPS
DP/$: $0.44 per DP GFLOPS

Without the DP, the novelty of the Titan wore off for me and i'm left with nothing but logic, which dictates there's no reason whatsoever so pay 2x the price of a 7970 GHz Edition for a GPU that only gets 400 more GFLOPS SP and less than 1/5 the DP. What are we actually paying for? Meh...

FWIW, the DP of the Titan *would* have been 1.5 TFLOPS if it wasn't crippled, blowing away the AMD 7970 GE.

Edit: As far as the 580 goes, it gets 197.6 GFLOPS DP - yes, even faster than the Titan and faster than the 680 which only has 128.8 GFLOPS DP, and the 580 costs less. You'll also get better SP compute results from the 580 than the 680.

[zombie67]

Thanks! That is exactly what I was looking for. Very clear now.

[Fire$torm]

So the 580 is the only CUDA card worth buying. The logic behind nVidia's marketing practices leaves me wondering how they managed to stay in business this long.

[DrPop]

So the 580 is the only CUDA card worth buying. The logic behind nVidia's marketing practices leaves me wondering how they managed to stay in business this long.

They aren't too worried about the crunchin' market, I'm sure. Either gamers (for things like the 680 or Titan) or of course "professionals" using their uber expensive, non- crippled "pro" teslas. Very, very lame for what we like to do with them.
I think the 570 is a good sweet spot card; the 580 is good, and of course the big dog double 590...but those are the last of the good Nvidia cards for crunching. Of course that's looking at it from a $/performance perspective like JPM was saying. And I don't see how to justify any other perspective on it!

[zombie67]

Edit: As far as the 580 goes, it gets 197.6 GFLOPS DP - yes, even faster than the Titan and faster than the 680 which only has 128.8 GFLOPS DP, and the 580 costs less. You'll also get better SP compute results from the 580 than the 680.

JPM: Maybe there is one more angle to look at, power consumption per GFLOPS?

I was reading this thread, and I wonder how titan compares to the previous generations?

The CC describes the GPU's abilities, and the researchers use this to determine the GPU type. The top CC2.0 GPU's (Geforce GTX 590, 580, 570, 480, 560 Ti (448), 470, 465) are less power efficient than the top CC3.0 GPU's (GTX690, GTX680, GTX670, GTX660Ti, GTX660, GTX650Ti, GTX650) making the CC3.0 cards preferable. The GTX570 and GTX580 are roughly as powerful as the GTX660Ti, GTX670 and GTX680, but being less power efficient and an older generation they are slightly down the recommended list.