PDA

View Full Version : No GTX680 until Q2 2012


Pages : [1] 2 3 4 5 6 7 8 9 10

Shamrock
02-07-12, 10:26 PM
More precisely, August/September.

Take it with a grain of salt. It's semi accurate.

http://semiaccurate.com/2012/02/07/gk110-tapes-out-at-last/

Just taped out a few weeks ago?

Rumors also going around the GTX690 will be $1000!

MUYA
02-07-12, 11:08 PM
Providing there are 1 or 2 revisions? Guess I will have to wait on that and be stuck with the same performance for another couple of months.

The jump from 580 GTX to 7970 does not seem a plausible return of investment for me. Now if the 7950's were to come down in price I would take the plunge...maybe for a cross fire solution.

Or if the Gk104 was that much faster than my 580 GTX i would be all over it too.

Some specs in some other website mention 4096 Sps!! I dunno about that....and maybe that's why it is expected to cost a kidney perhaps. But going from 512 to 4096 SPs (i am assuming linear scaling with everything normalized performance wise)...would be orgasmic....then again this is just rumors and not likely to be the case...(4096 SPs and stuff)

Viral
02-07-12, 11:40 PM
Romour has it that GK100 (original GTX680) was somewhat broken and failed to yield in any decent numbers so they went straight to the revision, GK110, so they could get it to market quicker by shifting focus away from the flawed part. It's like G100 vs G110, except this time the original 100 part is so bad they decided to skip it all together - or they just learned their lesson from last time?

The question is, is it better to delay the high end part 3-6 months to save face or is releasing a partially crippled, hot, power hungry, expensive chip in very short supply but on time the better option?

Viral
02-07-12, 11:42 PM
Some specs in some other website mention 4096 Sps!! I dunno about that....and maybe that's why it is expected to cost a kidney perhaps. But going from 512 to 4096 SPs (i am assuming linear scaling with everything normalized performance wise)...would be orgasmic....then again this is just rumors and not likely to be the case...(4096 SPs and stuff)

They wont be the same SP's and they won't have hot clocks. Think of them more in line with ATi's SP's, although likely somewhat less capable if there are 4096 of them. Unless the chip will be ~650mm^2. That would be.... interesting.

MUYA
02-08-12, 12:29 AM
They wont be the same SP's and they won't have hot clocks. Think of them more in line with ATi's SP's, although likely somewhat less capable if there are 4096 of them. Unless the chip will be ~650mm^2. That would be.... interesting.

you don't know that they are less capable...the way both camps design their SPs will have their benefits/caveats as well as workarounds :p comparable....but i wouldn't call it less capable. There are optimizations/ patented solutions to get the logic to do what they want which of course again may have certain conditions etc but u get my drift

Viral
02-08-12, 07:48 PM
If there's 4 times as much of them on a process that's only twice as small I'd say it's pretty safe to assume that they are less capable in some way. Even if it's just the lack of hot clocks - which makes a lot of sense, roughly half the clock speed but 4 times as many, so a doubling of potential throughput.

MUYA
02-08-12, 08:33 PM
If there's 4 times as much of them on a process that's only twice as small I'd say it's pretty safe to assume that they are less capable in some way. Even if it's just the lack of hot clocks - which makes a lot of sense, roughly half the clock speed but 4 times as many, so a doubling of potential throughput.

you are speculating in the negative order with even knowing performance yet. Then this is all conjecture either way at this point as we know bugger all. I tend to be cautious of the rumor that the GK110 is 4096 SPs.

fasedww
02-11-12, 01:42 PM
So the 680GTX won't be out tell Fall?:thumbdwn:

ViN86
02-11-12, 03:31 PM
you are speculating in the negative order with even knowing performance yet. Then this is all conjecture either way at this point as we know bugger all. I tend to be cautious of the rumor that the GK110 is 4096 SPs.

Viral is just speculating on what we will see if there are indeed 4096 SP's. I think he's spot on if the rumor is true. The thought of 4096 CUDA units is absurd, and doubling SP's while lowering clocks is a reasonable approach.

MUYA
02-13-12, 10:01 PM
Viral is just speculating on what we will see if there are indeed 4096 SP's. I think he's spot on if the rumor is true. The thought of 4096 CUDA units is absurd, and doubling SP's while lowering clocks is a reasonable approach.

This assumption is as much as mine, nothing to go but gut instincts perhaps (but we all aren't ASIC experts). Just because that the purported tactics for GK110 is to try to get more SPs in doesn't mean squat in terms of the capability of the SPs themselves. Is still am skeptical that the sp count could be that high!

If anything, some fervent hater of NV who runs a not so accurate website mentioned that the GK110 is/may be reticule limited which leads me (and only me perhaps) to believe that if it is 4096 SPs then, the SPs may not be so limited (then again it depends what market we are talking ab out I guess).

This shader thing talk is interesting. Its interesting in that the focus on DP related markets is taking it toll on both camps as that's the lucrative market (well maybe I sould say us desktop GAMERS! :p). Both these IHVs can't ignore that emerging market yet at the same time, both have to compromise to give something that is acceptable for HPC/professional market and also desktops.

LOL! i Love SILLY SEASON!!!

hmmm...what a load of codswallop! hahahah

MUYA
02-13-12, 10:03 PM
Man we should have these conversations in bar when we slightly pi$$ed! :p You knwo what i mean Redeemed? hahahahaha! :D

Redeemed
02-14-12, 01:45 AM
Uhh... Muya? :wtf:


:lol:

ViN86
02-14-12, 11:27 AM
This assumption is as much as mine, nothing to go but gut instincts perhaps (but we all aren't ASIC experts). Just because that the purported tactics for GK110 is to try to get more SPs in doesn't mean squat in terms of the capability of the SPs themselves. Is still am skeptical that the sp count could be that high!

If anything, some fervent hater of NV who runs a not so accurate website mentioned that the GK110 is/may be reticule limited which leads me (and only me perhaps) to believe that if it is 4096 SPs then, the SPs may not be so limited (then again it depends what market we are talking ab out I guess).

This shader thing talk is interesting. Its interesting in that the focus on DP related markets is taking it toll on both camps as that's the lucrative market (well maybe I sould say us desktop GAMERS! :p). Both these IHVs can't ignore that emerging market yet at the same time, both have to compromise to give something that is acceptable for HPC/professional market and also desktops.

LOL! i Love SILLY SEASON!!!

hmmm...what a load of codswallop! hahahah

Well, CUDA cores are much different from SP's. And if it is true that Nvidia is going the route of ATI, I am curious about the future of CUDA on consumer cards (i.e. non-workstation cards). There's no reason they can't run PhysX on SP's (might take a little driver rewrite), and most people don't use CUDA, so it may become a Quadro only feature.

Honestly, I hope that's not the case and we continue to get CUDA on consumer performance cards, but the architecture change worries me. Another option is to rework CUDA to be more like OpenCL. ATI cards outperform Nvidia cards in OpenCL applications currently. With some smart adaptation and architecture changes, Nvidia could change that around.

MUYA
02-14-12, 11:43 AM
By SPs..I actually meant NV's cuda core :p It probably is a bit too much for NV right now to design GPUs for different segments of the market....one core GPU with ...the fundamentally the same GPU shipped as different products ie quadros/ teslas and desktops with inhibitors in the drivers I guess.

ViN86
02-14-12, 11:57 AM
By SPs..I actually meant NV's cuda core :p It probably is a bit too much for NV right now to design GPUs for different segments of the market....one core GPU with ...the fundamentally the same GPU shipped as different products ie quadros/ teslas and desktops with inhibitors in the drivers I guess.

SP's are different from CUDA cores. Typically higher clocked and more efficient. That's why a 480 CUDA core GTX480 outperforms a 1600 SP HD5870.

EDIT: I wonder if Nvidia will change architecture but still refer to them as CUDA cores? :confused: Guess we won't know for sure until (hopefully) this week or next.

mullet
02-14-12, 01:40 PM
Waiting for this card is killing me.

Viral
02-14-12, 06:26 PM
you are speculating in the negative order with even knowing performance yet. Then this is all conjecture either way at this point as we know bugger all. I tend to be cautious of the rumor that the GK110 is 4096 SPs.

I'm just saying that the die size would be insanely large if they quadrouple the amount of the current CC's. Also, unless they beef up the rest of the chip accordingly and significantly increase the memory bandwidth those extra cores will go to waste due to imbalance in the architecture. If they're using that many cores, it's damn likely they will be different (weaker) than the current cores. I doubt they'll change the naming, however.

Assuming the cores wont be hot clocked, we can at least assume that they will only be half as capable at any given base clock speed. Surely the removal of hot-clock-ability will no doubt free up some space and complexity on the die, which in turn makes adding that many more cores possible.

Current rumour is that the GK104 will have 1536 CC's, think of that as a 768 CC GF1xx part. Of course clock speeds will be a bit more than half of the old hot clock speeds, most likely. I myself think they they will possibly be further reduced beyond just removing the hot clock, but not too drastically.

MUYA
02-14-12, 10:13 PM
SP's are different from CUDA cores. Typically higher clocked and more efficient. That's why a 480 CUDA core GTX480 outperforms a 1600 SP HD5870.

EDIT: I wonder if Nvidia will change architecture but still refer to them as CUDA cores? :confused: Guess we won't know for sure until (hopefully) this week or next.

It is given that both IHV classify a basic functional unit differently because of their adopted architecture. There is no standardization for this and we go by what these IHV own defined functional units are. Cuda cores for Nv and now GCN thing for ATI. Hence, the difference in the numbers of the count of these basic functional units and are directly not comparable is already a given.

My statement still is valid to whatever NV defines as their basic functional unit for Kepler. I doubt that they would fundamentally re-write the board with Kepler on this aspect (improved and optimized from Fermi ones though they may be). I think we were led to believe through different road maps that Maxwell is the architectural change and not Kepler.

With the process shrink from 40nm to 28nm this allows for more transistors to be packed in the same area but, from rumors the so called 1536 CC GK104 is 340mmsq or thereabouts. Given that someone running another website (who has a knack for getting industry news from moles) has stated that GK110 may be reticule limited, I may be inclined to follow that logic based on the Gk104 rumors that the 4XXX CC count for the rumored GK110 may be off. The actual CC may be lower...... Thus also my deduction that this basic function unit for Kepler series isn't suddenly all the inferior to Fermi ones.

I am guessing this makes sense when I am drunk :p

MUYA
02-14-12, 10:20 PM
I'm just saying that the die size would be insanely large if they quadrouple the amount of the current CC's. Also, unless they beef up the rest of the chip accordingly and significantly increase the memory bandwidth those extra cores will go to waste due to imbalance in the architecture. If they're using that many cores, it's damn likely they will be different (weaker) than the current cores. I doubt they'll change the naming, however.

Assuming the cores wont be hot clocked, we can at least assume that they will only be half as capable at any given base clock speed. Surely the removal of hot-clock-ability will no doubt free up some space and complexity on the die, which in turn makes adding that many more cores possible.

Current rumour is that the GK104 will have 1536 CC's, think of that as a 768 CC GF1xx part. Of course clock speeds will be a bit more than half of the old hot clock speeds, most likely. I myself think they they will possibly be further reduced beyond just removing the hot clock, but not too drastically.
FFS I will stop using SPS and use CC's as that's the way NV defines them :p

I am getting your point now. I think we are thinking along the same lines. You are extrapolating the fit of the GK110 from GK104 I guess or even current gen. ie if that was done it may be impossible etc :p So that is why you are inclined to think the CC may be of "lighter" weight if it is 4096 CCs (if this holds true). I tend to think the opposite and it will not be the 4096 CCs but the CCs (basic unit) will not be far different from existing fermi ones.

But then the purported changes to the hot clock thing could also mean changes to CCs...at least how they line up blah blah...best left to talking about that when you are drunk and you just nod your head as the world is spinning.....

..and i could be pulling this out of my arse as well hahahahahaha...all will be revealed i guess in a few days...PDXLAN or March

fasedww
02-15-12, 06:19 PM
Waiting for this card is killing me.
Whens it coming out?:)

mullet
02-16-12, 12:17 PM
I have a 5850 :(

i SPY
02-16-12, 01:12 PM
I'm just saying that the die size would be insanely large if they quadrouple the amount of the current CC's. Also, unless they beef up the rest of the chip accordingly and significantly increase the memory bandwidth those extra cores will go to waste due to imbalance in the architecture. If they're using that many cores, it's damn likely they will be different (weaker) than the current cores. I doubt they'll change the naming, however.

Assuming the cores wont be hot clocked, we can at least assume that they will only be half as capable at any given base clock speed. Surely the removal of hot-clock-ability will no doubt free up some space and complexity on the die, which in turn makes adding that many more cores possible.

[i]Current rumour is that the GK104 will have 1536 CC's, think of that as a 768 CC GF1xx part. Of course clock speeds will be a bit more than half of the old hot clock speeds, most likely. I myself think they they will possibly be further reduced beyond just removing the hot clock, but not too drastically.[/8]

yeah apparently its like so

I can tell you that the Mockup68.jpg is wrong.

The Cuda core numbers are correct but by this definition of Cuda core, you get 1 FMA per core versus 2 FMA as in Fermi.
So it's equivalent to a 768 Cuda core Fermi clock for clock in this regard.

Viral
02-16-12, 07:24 PM
yeah apparently its like so

Interesting, so if the hot clocks are gone, what's left is double the shader throughput, and whatever extra clock speed, memory clock/bus, cache improvements, ROP and texture unit increases. That's still over twice the potential performance of GTX 580, obviously without twice the memory bandwidth performance wont double, but it may exceed the bandwidth increase in many cases depending on what they do with cache and ROPs etc.

ViN86
02-16-12, 09:17 PM
yeah apparently its like so

That sounds right. Thanks for the info.

Muppet
02-17-12, 01:44 AM
I can wait. The 7970 is a nice upgrade now if I wanted it. But it is to expensive for the performance increase over the 580's.