GTX 980 Ti Potentially Launching Mid-Late May

[Commander Llama]

'Even a broken clock is right twice a day' might actually be correct - even in a probabilistic sense - with regards to WCCFTech's monthly video card speculation articles.

[Whorax]

980 ti is clocked higher so it should be faster why would you buy something because it is expensive

That doesn't have anything to do with anything.

[ErickS89]

Nvidia be like, you like our Titan? Well get ready for an all new beast, Nvidia GTX Titan (G)od!

[spartaman64]

That doesn't have anything to do with anything.

other than that the 980 ti will be faster

[MrRavens]

At the moment, I'm more compelled to get the "980Ti". But of course I still wanna see how AMD is going to pricing their new GPUs.

[Whorax]

other than that the 980 ti will be faster

It's becoming clear to me that you don't know how GPUs work.

[Prastupok]

It's becoming clear to me that you don't know how GPUs work.

 

I was going to say how business works.

[BiG StroOnZ]

My predictions are not a 980 Ti but a new series featuring GM200 main cards alongside Maxwell Refresh cards (like we saw with the 700 Series, three Cards featuring GK110 two cards featuring refresh of Kepler GK104 [then one much smaller card coming later in 2016 showing what Pascal has to offer like the 750 Ti did with Maxwell, so like a 1050 Ti]). If the rumors are this card is coming out in September, that would be exactly one year since Maxwell launched so it would put it in the same timeline as the 600 to 700 series.

 

Here are my estimated core counts:

 

GTX Titan X  - 3072 CUDA Cores 

GTX 1080 Ti / Ultra  - 2816 CUDA Cores (Increased Clock Speeds and Memory Speeds compared Titan X)

GTX 1080   - 2688 CUDA Cores (Increased Clock Speeds and Memory Speeds compared to Titan X

GTX 1070  - 2048 CUDA Cores (980 Refresh)

GTX 1060  - 1664 CUDA Cores (970 Refresh)

GTX 1050 Ti xxxx CUDA Cores (Pascal "testing the waters")

 

The reason I say this, going against the rumors of what all the websites are reporting is I imagine that there has to be many imperfect GM200 chips. Being a 601 mm² die, there has to be many defective units. Which means yields must not be so great, even though 28nm is mature, the die size alone will lead to imperfections. Which leads to me to believe they have many chips lying around that cannot be used as a Titan X. Which also would mean they cannot be used as a full GM200 core 980 Ti either. Which means they probably have to be cut down further with SMMs being disabled. And since Maxwell SMMs feature 128 shaders each, they could be disabled in increments of 128. Which leads me to believe that the gap between the Titan X and 1080 Ti / Ultra will be slightly greater than the gap between the 1080 Ti / Ultra and the 1080. But I imagine in order for them to remain competitive they will bump up the clocks on both the memory and core to further increase performance. Which means we can very well see a 1080 Ti / Ultra at $700 but practically nearing Titan X performance or matching it (clock speeds making up the difference in core counts) and custom AIB partner versions surpassing it with higher clock speed offerings and/or with custom PCBs (and of course not to forget after overclocked results). Then I see the 1080 as their $550 offering being maybe 10-13% slower than whatever the 1080 Ti / Ultra is. Then of course the 980 refresh drops to $350-400 like the GTX 770 was, and the 970 refresh drops to $250 like the GTX 760 was. 

[spartaman64]

It's becoming clear to me that you don't know how GPUs work.

i know how gpus work the 980 ti has a higher clock and it has the same architecture, cuda cores, rops  

[Whorax]

i know how gpus work the 980 ti has a higher clock and it has the same architecture, cuda cores, rops  

No, you don't, because based on your reasoning the GTX 980 should be faster than the Titan X.

[ErickS89]

snip

I think new naming scheme is going to happen, I mean it'll be the 8800 gt all over again.

[spartaman64]

No, you don't, because based on your reasoning the GTX 980 should be faster than the Titan X.

no the 980 doesnt have the same cuda cores and rops

[Whorax]

no the 980 doesnt have the same cuda cores and rops

You just said that higher clock speeds means a better GPU. Did you change your mind or do a quick Google to figure out how graphics cards work? Lol.

 

"980 ti is clocked higher so it should be faster why would you buy something because it is expensive"

[spartaman64]

You just said that higher clock speeds means a better GPU. Did you change your mind or do a quick Google to figure out how graphics cards work? Lol.

 

"980 ti is clocked higher so it should be faster why would you buy something because it is expensive"

no i said that because i assumed everyone knew the 980 ti has the same cuda cores etc because its right there in the original post i didnt say every gpu are faster because of higher clock i said the 980 ti is faster compared to titan x because of the higher clock

[patrickjp93]

Everything I can find shows the Titan X having 6.6 TFLOPS of single precision, and 0.2 TFLOPS of double precision. 

That's not in question here. What I'm saying is you can use the single-precision version of whatever algorithm it is you wish to implement and see how it scales to input size (memory usage) and core count and get a very good idea for what your peak double precision would be for a Tesla. They're great research tools.

[Whorax]

no i said that because i assumed everyone knew the 980 ti has the same cuda cores etc because its right there in the original post i didnt say every gpu are faster because of higher clock i said the 980 ti is faster compared to titan x because of the higher clock

I'm nearly positive they're not going to make a card that's faster than the card that is literally named after their top of the line GTX card.

 

You can think what you want but a 980Ti (assuming it ever happens) will not be faster than a Titan X.

[patrickjp93]

im not taking about pascal im talking about the current titan x and i assume you mean analogy not analogue because it would be strange if maxwell is an analogue processor and i dont quite understand your last sentence but all the current maxwell cards doesnt have good DP performance and i dont think you can use SP to substitute for DP

No, look at the definition of the word analogue as a noun. And yes you can, because DP will only ever be up to 1/2 the performance of SP versions of algorithms.

[spartaman64]

No, look at the definition of the word analogue as a noun. And yes you can, because DP will only ever be up to 1/2 the performance of SP versions of algorithms.

oh i thought you are talking about this kind of analogue http://en.wikipedia.org/wiki/Analogue_electronicslol and i dont think using single precision to replace double precision will work because a int in programming stores much less data than a double

[Rock48]

int in programming stores much less data than a double

​while that is true, you shouldn't be comparing integers and double-precision floating point values; they store totally different data.

int = float = 32-bit < long = double = 64-bit

[patrickjp93]

​while that is true, you shouldn't be comparing integers and double-precision floating point values; they store totally different data.

int = float = 32-bit < long = double = 64-bit

And even then the algorithms for add/sub/mul/div for ints and floats are very different and take different clock cycle counts.

[Cookybiscuit]

I was happy with my 780 until I played GTA V.  <_<

[GrimNeo]

If it's just a Titan X with 6GBs of vram and a higher base clock then I'm going with the Titan X because overkill . Just waiting for 2 of them to come back in stock from EVGA.

[zacRupnow]

It's for people who want the best most expensive single gpu in the world.

Until next year...

[BarisERY]

I dont think this will happen till amd launches its 390X because otherwise it diesnt make sense it will just cripple titan x sales.

[1Scotty1]

I am perfectly happy with my 970 SLI... some cards are overpriced for what they are - 980 for example