Any GeForce 660/660 Ti Speed Theories?

[Phynaz]

Also, Nvidia really did focus hard on perf/W this gen.
Everything in design was subjected to improving Perf/W.

This. Low power, high performance.

Oh, and CUDA.

 

[The_Golden_Man]

Oh, and I think my GTX 570 started showing signs of failing even before the release of the GTX 680. So had the GTX 570 died in that period I very likely could have ended up with a 7950 or something.

 

[RussianSensation]

Oh, and I think my GTX 570 started showing signs of failing even before the release of the GTX 680. So had the GTX 570 died in that period I very likely could have ended up with a 7950 or something.

Did you at least get the failed card covered by warranty? I hope so!

 

[The_Golden_Man]

Did you at least get the failed card covered by warranty? I hope so!

I did not try, because that card was bought through my company. And when bought through a company the warranty is only 1 year. That's the rules, at least here in Norway. So it went in the garbidge. Of course I tried the oven trick, but I noticed afterwards that one of the Powerphases had burnt out anyway.

 

[RussianSensation]

Apparently rumours are swirling that there are 2 separate GTX660 cards:

1344 SP / 192-bit memory bandwidth ~ 2,000 3dMark11
1152 SP / 256-bit memory bandwidth ~ 2,700 3dMark11

The 192-bit memory bandwidth seems to become a huge liability despite 1344 SPs.

 

[The_Golden_Man]

Apparently rumours are swirling that there are 2 separate GTX660 cards:

1344 SP / 192-bit memory bandwidth ~ 2,000 3dMark11
1152 SP / 256-bit memory bandwidth ~ 2,700 3dMark11

The 192-bit memory bandwidth seems to become a huge liability despite 1344 SPs.

The fact that 192Bit Mem Bandwidth becomes a huge bottleneck for this card does not surprise me one bit. Reason being both the GTX 670 and 680 scales very well with increased memfrequency.

 

[ShintaiDK]

Doing a random guess, I would say its 256bit. Despite the 550TI being 192bit with 2GB. But 550TI was a true 192bit design.

The memory controller is pretty small. And I guess the chance of having 2 defect clusters is higher than having a defect memory controller. And memory controller and clusters aint linked. Not to mention that each GPU would basicly have to be traced to what memory controller part was defect. That just sounds like abit of a mess.

Maybe 5.5Ghz memory as well to save some cash too in production.

 

[RussianSensation]

Fake?

24 ROPs
192-bit bus
1152
1.5GB of VRAM.

Jeez, it better be fake. With those specs, the card will be lucky to beat an HD7870.

 

[Atreidin]

I find it funny to read people so confident about their random guesses, pretending Nvidia is beholden to some pattern, real or imagined.

 

[ShintaiDK]

Little problem tho. The driver doesnt support 660TI. There are device entries for 690, 680 and 670. But no ID in the same family for a 660TI.

So unless they modded the drivers somehow. They dont support 660TI.

Also GPU-Z would have to know 660TI specs to fill in the info. It doesnt.

So I say its a fake SS. Info is right or not is a different matter.

 

[RussianSensation]

I find it funny to read people so confident about their random guesses, pretending Nvidia is beholden to some pattern, real or imagined.

No one here is confident about the accuracy of the specifications, but based on the rumored specifications (192-bit vs. 256-bit memory bandwidth, 1.5GB vs. 2GB of VRAM), we can make educated guesses as to how the card will perform. Considering GTX660Ti's main competition is HD7950 that at 1.05 ghz = GTX680, you can see that GTX660Ti has its work cut out for it. So it pretty much needs 32 ROPs, 2GB and 256-Bit to even be a contender. This is how threads like this usually are. People take wild stabs at specs/performance, hence the thread title "Theories". You don't have to participate if this is of no interest to you.

 

[Crap Daddy]

Fake?

24 ROPs
192-bit bus
1152
1.5GB of VRAM.

Jeez, it better be fake. With those specs, the card will be lucky to beat an HD7870.

What if it's the 660 non Ti? At 250$?

 

[Atreidin]

No one here is confident about the accuracy of the specifications, but based on the rumored specifications (192-bit vs. 256-bit memory bandwidth, 1.5GB vs. 2GB of VRAM), we can make educated guesses as to how the card will perform. Considering GTX660Ti's main competition is HD7950 that at 1.05 ghz = GTX680, you can see that GTX660Ti has its work cut out for it. So it pretty much needs 32 ROPs, 2GB and 256-Bit to even be a contender. This is how threads like this usually are. People take wild stabs at specs/performance, hence the thread title "Theories". You don't have to participate if this is of no interest to you.

I'm just commenting that I enjoy reading angsty comments where people get angry about people disagreeing with them, when they are still giving wild guesses based on patterns they pretend Nvidia has to hold to. There is no science to this, it is all guessing without direct knowledge of Nvidia's business plan. Good-natured discussion is fun but we are losing that, if we ever had it.

I prefer a good discussion full of people not taking things so seriously, but this place has had more and more veiled hostility and angst. I'm just saying it is still fun to read those posts, imagining the rage behind them. Maybe preempting it and pointing that out would derail some of it, I don't know. Usually these threads follow a pattern of innocuous question, polite discussion, tense discussion, ANGST AND RAGE, as (usually a few) people get worked up over stupid things.

 

[RussianSensation]

What if it's the 660 non Ti? At 250$?

Ya that would make more sense. Still not that great since HD7850/7870 cards launched March 5th. Nearly 6 months later for similar performance from a $250 GTX660 is not really adding any more bang-for-the buck. I was hoping for GTX660 at $199-229 with 7870 level of performance and GTX660Ti at $279-289 with HD7950 800mhz level OR $299 for GTX660Ti ~ HD7950 950mhz level. That would put price pressure on AMD and bring faster performance at lower price levels. I think NV may play the power consumption/efficiency card. If 660Ti is 5-7% faster performance than the 800mhz 7950 and uses 20-30W less power, people who don't overclock will buy it for sure. However, dropping power consumption from 7950's levels is going to be very tough card since 7950 consumes ~ 144-145W of power at load. Even in factory preoverclocked 860-900mhz states, it's still very frugal.

 

[ShintaiDK]

Here is another fake from the same person btw.

 

[Crap Daddy]

If they are not fake? 192bit and 24 ROPs for the 660 and the 256bit bus for the 660Ti. The same GK104. There's no GK106.

 

[Silverforce11]

If they are not fake? 192bit and 24 ROPs for the 660 and the 256bit bus for the 660Ti. The same GK104. There's no GK106.

Spot on.

 

[ShintaiDK]

If they are not fake? 192bit and 24 ROPs for the 660 and the 256bit bus for the 660Ti. The same GK104. There's no GK106.

They are fake.

 

[tviceman]

Seriously, Nvidia's time to get GK106 to market is really, really strange. Not because it is lagging behind AMD's 7000 series lineup, but because it is lagging behind it's own family of GPU's. So weird.

Anyways, I've said several times that Nvidia would be best to release a 256-bit 6smx card as well as a 192-bit 6smx card (both based on GK104). Of course their quantities and yields would dictate whether they could do that or not, but I don't think they would have enough to supply over the course of the 600 series product life unless they started using perfectly good (or better than needed) chips to fill these products. So without GK106, and if GK106 is in fact only half of GK104, they have put themselves between a rock and a hard place for the sweet spot GPU.

 

[RussianSensation]

From Dell Alienware X51 configuration (European website):

Looks like 1.5GB of VRAM for the non-Ti version is confirmed.

 

[The_Golden_Man]

From Dell Alienware X51 configuration (European website):

Looks like 1.5GB of VRAM for the non-Ti version is confirmed.

Is it likely the Non-Ti version will have 192Bit membandwith because of 1.5GB VRam then?

 

[Jaydip]

192Bit is buswidth not bw

 

[RussianSensation]

Is it likely the Non-Ti version will have 192Bit membandwith because of 1.5GB VRam then?

Ya, crippled on both sides sounds like non-Ti. Probably meant to be an HD7870 competitor.

NV is also moving to stop overclocking on GTX600 series line by denying warranty claims from AIBs. That would make it basically impossible for GTX660Ti to be better than the HD7950 + 1.1ghz OC if AIBs get nervous and tell the consumers that they won't cover their high overclocks. Interesting development that NV is going so far out of its way to stop overclocking on Kepler.

 

[The_Golden_Man]

192Bit is buswidth not bw

Sorry, I meant Buswidth

 

[SirPauly]

Seriously, Nvidia's time to get GK106 to market is really, really strange..

I don't think so based on how they officially offered constrained and need more 28nm wafers. They may be targeting the cores that make the most sense to get out first -- and when constrains lessen -- add more to the 28nm family.