GTX 670 NDA dropping at Midnight?

[Stuka87]

Kepler has outsold 7970 by 9:1 ratio in April according to Steam, and is readily avilable in the whole wide world, except in USA.
Where apparently Nvidia is not picking the money of the table
(but neither are etailers, like in EU where 7970 is €100 cheaper)
and hence demand far outweighs supply.

I want a link showing that 680's outsold 7970 by a 9:1 ratio. Because looking at Steams stats, I don't even see the GTX 680 listed for April at all.

And the reason it is readily available in the rest of the world is because it is priced FAR higher than a 7970. In some cases over $150 more than a 7970.

 

[blastingcap]

That's a nice set of pics.
But what are they supposed to prove?



Kepler has outsold 7970 by 9:1 ratio in April according to Steam, and is readily avilable in the whole wide world, except in USA.
Where apparently Nvidia is not picking the money of the table
(but neither are etailers, like in EU where 7970 is €100 cheaper)
and hence demand far outweighs supply.

Hey fish what graphics card(s) do you use in your fastest computer system?

 

[LOL_Wut_Axel]

That's a nice set of pics.
But what are they supposed to prove?



Kepler has outsold 7970 by 9:1 ratio in April according to Steam, and is readily avilable in the whole wide world, except in USA.
Where apparently Nvidia is not picking the money of the table
(but neither are etailers, like in EU where 7970 is €100 cheaper)
and hence demand far outweighs supply.

So where is this impressive statistic from Steam? Care to post a picture here? All I'm smelling right now is BS.

AMD is selling a lot more 28nm GPUs than NVIDIA right now.

 

[RampantAndroid]

I want a link showing that 680's outsold 7970 by a 9:1 ratio. Because looking at Steams stats, I don't even see the GTX 680 listed for April at all.

And the reason it is readily available in the rest of the world is because it is priced FAR higher than a 7970. In some cases over $150 more than a 7970.

Their stats are messed up. Look under DX10 and DX11 systems, not under GPUs. In both, you see the 7970's market share jumped by .01%. The GTX680's share jumped by .09%.

http://store.steampowered.com/hwsurvey?platform=pc

If in May the 680 and 7970 sales continue along that trend, the 680 will overtake the 7970. At least, as far as Steam is concerned.

 

[Ariae]

cool story brah

Please don't thread crap in the technical forums.
-ViRGE

 

[RampantAndroid]

cool story brah

Nice post edit. Thank you for not adding anything to this thread.

 

[Ariae]

Nice post edit. Thank you for not adding anything to this thread.

You're very welcome

oh lets see if you can catch this one

I still don't see how he's getting the 9:1 ratio by just looking at the market share increase? Someone wanna explain?

 

[RussianSensation]

Yeah it's usually about ~110W at 1300mv for the 7970. I agree with you on the gk104. They definitely could have made a ~275W board and given us voltage control up to ~1300-1400mv. Then people would be seeing 1400/1800mhz clocks on those bad boys doing exactly like you said +60% over the 580 instead of +40%

NV should have at least allowed AIBs to ship after market cards with beefier VRMs, cooling and unlocked voltage. With just 1.212V, 680 seems to hit 1350-1400mhz. I don't understand why NV had to lock out the voltage completely. After allowing EVGA to release 850mhz GTX460 FTW edition, this seems almost backwards.

Perhaps NV is going to launch "K" series chips with unlocked voltages. Here you go for $50 more GTX680-K with fully unlocked voltage for overclocking -- GPU overclocking of the future, the Intel way. :hmm:

Kepler doesn't scale nearly as well as Tahiti with overclocking, though.

You are confusing overclocking headroom with GPU performance scaling via additional frequency. Kepler scales fine, but it has low overclocking headroom.

GTX680 operates at a minimum of 1058mhz, with most cards hitting 1110mhz. With overclocking a lot of cards max out at just 1230-1250mhz. Therefore, the GPU scaling should be commensurate with the frequency increase (which is 1250 / 1110 = 13%). In fact, a lot of times GTX680 goes above 1110mhz, which means with average overclocking, GTX680's performance should not grow by more than 10-13% (and benchmarks are showing this because it's exactly what's expected).

OTOH, Tahiti overclocks from 925mhz to let's say 1200mhz (or a 30% overclock). In other words, it's not scaling that's the problem with GTX680 or that HD7970 GPU has some magical GPU scaling, but that most of the overclocking headroom has been "eaten away" by dynamic GPU Boost on the 680. It's incorrect to state that GTX680 has poor overclocking scaling.

Poor GPU scaling with overclocking means if you increase GPU frequency by 30%, GPU performance only grows by a fraction of the GPU frequency increase. That is not the case with Kepler.

 

[RampantAndroid]

NV should have at least allowed AIBs to ship after market cards with beefier VRMs, cooling and unlocked voltage. With just 1.212V, 680 seems to hit 1350-1400mhz. I don't understand why NV had to lock out the voltage completely. After allowing EVGA to release 850mhz GTX460 FTW edition, this seems almost backwards.

Perhaps NV is going to launch "K" series chips with unlocked voltages. Here you go for $50 more GTX680-K with fully unlocked voltage for overclocking. :hmm:

Remember the 590s that cooked when the voltage was increased? Might be an answer to why.

 

[LOL_Wut_Axel]

NV should have at least allowed AIBs to ship after market cards with beefier VRMs, cooling and unlocked voltage. With just 1.212V, 680 seems to hit 1350-1400mhz. I don't understand why NV had to lock out the voltage completely. After allowing EVGA to release 850mhz GTX460 FTW edition, this seems almost backwards.

Perhaps NV is going to launch "K" series chips with unlocked voltages. Here you go for $50 more GTX680-K with fully unlocked voltage for overclocking -- GPU overclocking of the future, the Intel way. :hmm:


You are confusing overclocking headroom with GPU frequency scaling. Kepler scales fine, but it has low overclocking headroom.

GTX680 operates at a minimum of 1058mhz, with most cards hitting 1110mhz. With overclocking a lot of cards max out at just 1230-1250mhz. Therefore, the GPU scaling should be commensurate with the frequency increase (which is 1250 / 1110 = 13%). OTOH, Tahiti overclocks from 925mhz to let's say 1200mhz (or a 30% overclock).

In other words, it's not scaling that's the problem with GTX680, but that most of the overclocking headroom has been "eaten away" by dynamic GPU Boost. It's incorrect to state that GTX680 has poor overclocking scaling.

No, it has small overclocking headroom and you don't gain as much as Tahiti for every percent you clock it higher which, again, isn't that big of an issue because the HD 7970 will "only" catch up.

 

[RussianSensation]

Remember the 590s that cooked when the voltage was increased? Might be an answer to why.

I think that's because NV didn't allow AIBs to change the components on the 590.

- The VRMs can be increased
- The power can be raised to 300W with dual 8-pins
- The cooling can be improved.

All that has been done:

And they still locked the BIOS...:thumbsdown:

 

[RussianSensation]

Single slot GTX670....maybe

GTX670 = tiny PCB = mid-range GPU.

Some speculation that memory clocks will be lowered from 6008 to ~5000, dropping memory bandwidth from 192 GB/sec to ~160 GB/sec. Weak sauce.

 

[3DVagabond]

NV should have at least allowed AIBs to ship after market cards with beefier VRMs, cooling and unlocked voltage. With just 1.212V, 680 seems to hit 1350-1400mhz. I don't understand why NV had to lock out the voltage completely. After allowing EVGA to release 850mhz GTX460 FTW edition, this seems almost backwards.

Perhaps NV is going to launch "K" series chips with unlocked voltages. Here you go for $50 more GTX680-K with fully unlocked voltage for overclocking -- GPU overclocking of the future, the Intel way. :hmm:

You are assuming they have any GK104 chips to sell the AIB's to make these "unlocked" cards.

The 460 FTW served it's purpose. If the same situation arrives again, we might see the same thing happen.

The "K" series is an interesting thought. :nodshead:

 

[blastingcap]

Snip

I have no idea where you find this stuff, lol. Anyway, it would seem that the gtx680 was supposed to be the gtx670ti. So this new card that is being called the gtx670 in rumors... should have been the gtx660.

Why does this card have such a big shroud (and presumably heatsink) relative to the PCB?

My speculation: Maybe it wasn't supposed to, but since Nvidia needed an answer to the 7950, this little card was conscripted for the task. Only problem was that it needed high clocks to beat the 7950 (stock vs. stock), which put off a lot of heat, so in the end the little card had to get an oversized cooling assembly to keep heat and noise tolerable.

 

[OCGuy]

nV has silenced a lot of critics since Fermi 1. Bravo.

 

[Don Karnage]

NV should have at least allowed AIBs to ship after market cards with beefier VRMs, cooling and unlocked voltage. With just 1.212V, 680 seems to hit 1350-1400mhz. I don't understand why NV had to lock out the voltage completely. After allowing EVGA to release 850mhz GTX460 FTW edition, this seems almost backwards.

Perhaps NV is going to launch "K" series chips with unlocked voltages. Here you go for $50 more GTX680-K with fully unlocked voltage for overclocking -- GPU overclocking of the future, the Intel way. :hmm:



You are confusing overclocking headroom with GPU performance scaling via additional frequency. Kepler scales fine, but it has low overclocking headroom.

GTX680 operates at a minimum of 1058mhz, with most cards hitting 1110mhz. With overclocking a lot of cards max out at just 1230-1250mhz. Therefore, the GPU scaling should be commensurate with the frequency increase (which is 1250 / 1110 = 13%). In fact, a lot of times GTX680 goes above 1110mhz, which means with average overclocking, GTX680's performance should not grow by more than 10-13% (and benchmarks are showing this because it's exactly what's expected).

OTOH, Tahiti overclocks from 925mhz to let's say 1200mhz (or a 30% overclock). In other words, it's not scaling that's the problem with GTX680 or that HD7970 GPU has some magical GPU scaling, but that most of the overclocking headroom has been "eaten away" by dynamic GPU Boost on the 680. It's incorrect to state that GTX680 has poor overclocking scaling.

Poor GPU scaling with overclocking means if you increase GPU frequency by 30%, GPU performance only grows by a fraction of the GPU frequency increase. That is not the case with Kepler.

Nv has effectively killed overclocking with Kepler.

 

[toyota]

I have no idea where you find this stuff, lol. Anyway, it would seem that the gtx680 was supposed to be the gtx670ti. So this new card that is being called the gtx670 in rumors... should have been the gtx660.

Why does this card have such a big shroud (and presumably heatsink) relative to the PCB?

My speculation: Maybe it wasn't supposed to, but since Nvidia needed an answer to the 7950, this little card was conscripted for the task. Only problem was that it needed high clocks to beat the 7950 (stock vs. stock), which put off a lot of heat, so in the end the little card had to get an oversized cooling assembly to keep heat and noise tolerable.

the gk104 was almost certainly originally planned for gtx660ti. they probably started calling it the gtx670ti because they saw that it was being fairly competitive between the 7950 and 7970. they probably bumped the clocks and decided it was good enough to be called the gtx680. I just dont think it was ever intended to be anything above gtx660ti when originally conceptualized. everything about its architecture shows that to be logical. also a gtx670ti would have likely been a cut down gtx680 if a big gtx680 would have been released.

 

[Zebo]

the gk104 was almost certainly originally planned for gtx660ti. they probably started calling it the gtx670ti because they saw that it was being fairly competitive between the 7950 and 7970. they probably bumped the clocks and decided it was good enough to be called the gtx680. I just dont think it was ever intended to be anything above gtx660ti when originally conceptualized. everything about its architecture shows that to be logical. also a gtx670ti would have likely been a cut down gtx680 if a big gtx680 would have been released.

That sounds about right. But it's a fantastic card that's better than last gens top end and AMDs this gen. $500 just made business sense.

I think a lot of people are holding out for real Kepler though. GTX 780.

 

[RussianSensation]

"Huang said that by the time of release the company will have invested about $2 billion in the design. Little information was given on performance, but Huang was adamant they there were targeting performance-per-watt, and that is what the scale on the left hand side represents. Kelper will have apparently have 4x the performance per watt in terms of double precision GFLOPS." ~ Source

Peak power consumption

GTX480 = 272 W
GTX580 = 229 W
GTX680 = 186 W

Double Precision performance (1/8 of SP for Fermi architecture and 1/24th of SP for Kepler GK104 derivative):

GTX480 = 168 GFLOPS
GTX580 = 198 GFLOPS
GTX680 = 135 GFLOPS (1536 SP x 2 Ops/clock x 1058mhz GPU Boost x 1/24th Single precision / 1000)

Double Precision performance / Watt:

GTX480 = 0.62 GFLOPS /W
GTX580 = 0.86 GFLOPS / W
GTX680 = 0.73 GFLOPS / W (worse than full harvested Fermi GF110) :sneaky:

Nvidia's Fermi - Kepler - Maxwell Slide does NOT compute.

GK104 = mid-range

 

[RussianSensation]

You are assuming they have any GK104 chips to sell the AIB's to make these "unlocked" cards.

They don't need to unlock the card. NV can let them unlock the voltages via custom BIOSes. AIBs can make 670/680s with beefier VRMs, 2x8-pin connectors and improved cooling. 1.25V can be pushed into 7970 on 28nm. Surely we can pump 1.25V into GK104 to clock it to 1400mhz if cooling permits?

 

[blastingcap]

snip

They stripped a lot of HPC functionality out of gk104, so what do you expect? Of course gtx680 wasn't intended to be the flagship Kepler.

 

[toyota]

They stripped a lot of HPC functionality out of gk104, so what do you expect? Of course gtx680 wasn't intended to be the flagship Kepler.

they cant strip what it never really had.

 

[blastingcap]

they cant strip what it never really had.

You know what I mean. Gtx680's heritage is suspect and it looks like the successor to gtx560, not gtx580. A true successor to gtx580 would probably have tons better DP performance.

 

[toyota]

You know what I mean. Gtx680's heritage is suspect and it looks like the successor to gtx560, not gtx580. A true successor to gtx580 would probably have tons better DP performance.

yes I know and was just kidding with you on the way you worded it. and you know I agree that the gk104 was originally planned to be a gtx660ti type card.

 

[3DVagabond]

They don't need to unlock the card. NV can let them unlock the voltages via custom BIOSes. AIBs can make 670/680s with beefier VRMs, 2x8-pin connectors and improved cooling. 1.25V can be pushed into 7970 on 28nm. Surely we can pump 1.25V into GK104 to clock it to 1400mhz if cooling permits?

They would need to make cards with beefier a PCB. Then they would need gpu's to put on them. If all of the gpu's are going on the reference boards, they can't make the custom cards.