NVIDIA Plans Two New SKUs Based on GK106

[btarunr]

A little later this month, NVIDIA plans to launch the GeForce GTX 660, its first desktop GPU based on the 28 nm GK106 silicon. The GTX 660 carries NVIDIA's ASIC codename GK106-400. An internal document detailing other SKUs based on the GK106 in the works, was leaked to the web. According to a 3DCenter.org report, the two ASIC codenames GK106-250 and GK106-200 described in the document could be new SKUs.

The GK106-250, carrying a working title GeForce GTX 655, could have one of the five SMX units of the GK106 disabled, resulting in a CUDA core count of 768. According to the document, it features a 192-bit wide GDDR5 memory interface, holding 2 GB of memory, much like the GTX 660. The GK106-200, on the other hand, with a working title GeForce GTX 650 Ti, could feature just three of the five SMX units on the GK106 silicon, making up 576 CUDA cores. According to the document, it features 2 GB of GDDR5 memory across a 128-bit wide memory interface, much like the GK107-based GTX 650. The two SKUs could help NVIDIA seal gaps in its mainstream desktop GPU lineup.

View at TechPowerUp Main Site

 

[dj-electric]

Now that will fill the weird gap...

 

[Harlequin_uk]

7 months after AMD stamped all over the mid range.....

 

[hardcore_gamer]

Tweaktown already reviewed 660.

 

[HumanSmoke]

Tweaktown already reviewed 660.

...with recycled benchmark scores on older drivers for most cards, on a bevy of outdated games. Very helpful [/sarcasm]
You're a little late anyhow, the TT article has already been done to death- surprisingly enough in the GTX 660 news articles (namely here and here)

 

[damric]

What I get out of this:

GK106 yields bad.

GDDR5 cheap.

 

[hardcore_gamer]

...with recycled benchmark scores on older drivers for most cards, on a bevy of outdated games. Very helpful [/sarcasm]

They did the same with 670.

"The GEFORCE GTX 670 is just an awesome video card that is going to be a serious headache for AMD. The performance of the video card is really quite crazy and the fact that this is the second option for NVIDIA when it comes to single GPU video cards is bad for AMD when it has no issue competing with their strongest option; the HD 7970"

Their results were later confirmed by other review sites when NDA lifted. I consider it as "helpful". I expect the same with 660 review.

Could you please point me to a better gtx660 review right now ? Thanks in advance.

 

[SIGSEGV]

where is GTX670ti anyone? Nvidia's Ti naming make me laugh...
why the hell on earth they make a such weird names like that.. *Le me sigh

 

[Harlequin_uk]

there allready is a GTX 670Ti - its called the GTX 680

 

[HumanSmoke]

They did the same with 670.
Their results were later confirmed by other review sites when NDA lifted. I consider it as "helpful". I expect the same with 660 review

TT's review was somewhat more favourable than most other review sites -largely due to game selection. TT also did the same with the 660 Ti - and according to their tests the 660Ti bests the HD 7950 in 65% of benchmarks, and of ~ half the benchmarks it "loses", it does so by a solitary frame per second. Most other sites have the 660Ti and 7950 pretty much even at 1920x1080/1200 with the 7950 pulling ahead at 2560x1440/1600 (including W1zz's TPU benches I might add)

Could you please point me to a better gtx660 review right now ? Thanks in advance.

That's kind of the point. TT jumps the gun pretty much pissing on every other sites chips...then buys page clicks by bleating on about Nvidia's bias towards them, while trying to make a case for Official sample = NDA, while obtaining a sample under the counter =/= NDA.
I doubt you'll find many supporters of this site, or any other enthusiast site (TT aside) that are happy to publicize/condone TT's methodology tbh....but each to their own.

 

[Harlequin_uk]

Thing is though - TT have been pretty much spot on for the last 4 or 5 cards they have released `early` , so simply saying `old drievrs or old games` doesnt quite cut it - you can say the same if you like for tomshardware or anand or H or any of them that review equipment; they dont retest older cards with the latest drivers each and every time something new and shiny comes along - time is the factor , when it takes 3 or 4 hours at least to get a good and average set of benchmark reuslts in.

 

[alwayssts]

They did the same with 670.

"The GEFORCE GTX 670 is just an awesome video card that is going to be a serious headache for AMD. The performance of the video card is really quite crazy and the fact that this is the second option for NVIDIA when it comes to single GPU video cards is bad for AMD when it has no issue competing with their strongest option; the HD 7970"

Their results were later confirmed by other review sites when NDA lifted. I consider it as "helpful". I expect the same with 660 review.

Could you please point me to a better gtx660 review right now ? Thanks in advance.

First, read this review or even just this page...Hopefully you will see how sad marketing makes me.

http://hardocp.com/article/2012/08/23/galaxy_gtx_660_ti_gc_oc_vs_670_hd_7950/3

TLDR:

The point is that 1568 (1344 + 224 sfu)/1792 equals 87.5%...EXACTLY THE SAME AS 1120/1280 (660 vs 7870). Does a 670 compete with 7950 at similar clocks? This is fairly clearly shown to be not the case.

1195/1316 = 90.8%. That means the radeon should be be 3-4% faster on paper if limiting factor is the shader core. In reality, it is 2-3% faster, within a percent or two I would call margin of error. The 7870 has no tdp/clock limitations, while 660 very well may (150w tdp, bigger die, faster memory granted on a smaller bus). Even if it doesn't, the ipc makes it immaterial. Have I beaten this horse dead-enough yet?

I would like to note 1300mhz is not a typical overclock (you can look at TPUs results across a gamut of cards...or the gpureview.com sample across all launch reviews), and 1195 is. Tdp/powertune is 200w for both cards, but the greater memory (probably ~20w) may often hold the core clock back on 7950 vs 670. Bandwidth is immaterial for the 7950 because of the wider bus, even for 670 the bandwidth (some cards will clock higher)the overclock is sufficient to feed the shaders because (afaict given performance) the sfus use the internal cache. GK104 is natively very efficiently well-designed for the clocks of high-end memory vs. the 28nm process and 225w. I still fail to see how that efficiency is worth 100-200$ more...although one could quibble on card form-factor/power consumption/noise are.

Again, there is a ROP bottleneck slightly above 1792 shaders (7950/680)...or roughly >6SMX or 20 CUs/24 (but really 32 for AMD) ROPs when talking practical configurations. 7970 is a pretty value-less card unless you are doing gpgpu or need the small usable ipc difference (~3%) and higher tdp allowance for clock potential similar to the 670/680 (which may net you a another 5%). It has 2048 shaders and the wide bus because when it is clocked to >1100mhz 2048 shaders can use more than 256-bit/7gbps for compute, while also securing it will always be the fastest 32 ROP part per clock. It also doubles as a buffer for getting acceptable yields on a 1792sp part (the most efficient)...unlike gk104 which does that natively, hence we get less-efficient (core) salvage parts and this clockspeed game and nvidia charges you a pretty penny for the parts that all work.

Point being, you have the right to believe what you want and that is fine...but in this case it makes you look ill-informed...or exactly what nvidia thrives on...and I continue to wish people knew better.

Anyway...655 and 650 = probably stock-clocked to equal 7850 and 7770. 650 should be better than 7770, if barely, and 655 is a wildcard only because 7850 is such a funky product (some have low voltage and no voltage control, others bios locked at 1050mhz etc while others have none of those quirks...655 @ 1200 [average for 1.175v] = 7850 @ 1050). I stand by that whatever replaces the 7700 will be clocked to beat these parts; the top model will probably have similar ipc to 655 with higher stock clock, the lesser model more ipc than 650 and a lower clock. I still think ~-900-925mhz and ~1075-1100mhz will be the pro and xt spots going up the lineup for 8000, which would line up well against these cards.

 

[dj-electric]

Tweaktown already reviewed 660.

lol.
TT's reviews are pathetic.
Old drivers, pathetic set of benchamrks - old games, old versions of them
No overclock, no overclocked performance really. No power consumption or noise.
No super high resolutions, no multimonitor setups, and the list goes on and on.
It is funny how the site with the worst reviews gets them first
it makes us, reviewers feel good about holding and releasing the "wow" effect of users responsing to the review and it's results. No worries when it comes to TT.

 

[mlee49]

First, read this review or even just this page...Hopefully you will see how sad marketing makes me.

http://hardocp.com/article/2012/08/23/galaxy_gtx_660_ti_gc_oc_vs_670_hd_7950/3

TLDR:

The point is that 1568 (1344 + 224 sfu)/1792 equals 87.5%...EXACTLY THE SAME AS 1120/1280 (660 vs 7870). Does a 670 compete with 7950 at similar clocks? This is fairly clearly shown to be not the case.

1195/1316 = 90.8%. That means the radeon should be be 3-4% faster on paper if limiting factor is the shader core. In reality, it is 2-3% faster, within a percent or two I would call margin of error. The 7870 has no tdp/clock limitations, while 660 very well may (150w tdp, bigger die, faster memory granted on a smaller bus). Even if it doesn't, the ipc makes it immaterial. Have I beaten this horse dead-enough yet?

I would like to note 1300mhz is not a typical overclock (you can look at TPUs results across a gamut of cards...or the gpureview.com sample across all launch reviews), and 1195 is. Tdp/powertune is 200w for both cards, but the greater memory (probably ~20w) may often hold the core clock back on 7950 vs 670. Bandwidth is immaterial for the 7950 because of the wider bus, even for 670 the bandwidth (some cards will clock higher)the overclock is sufficient to feed the shaders because (afaict given performance) the sfus use the internal cache. GK104 is natively very efficiently well-designed for the clocks of high-end memory vs. the 28nm process and 225w. I still fail to see how that efficiency is worth 100-200$ more...although one could quibble on card form-factor/power consumption/noise are.

Again, there is a ROP bottleneck slightly above 1792 shaders (7950/680)...or roughly >6SMX or 20 CUs/24 (but really 32 for AMD) ROPs when talking practical configurations. 7970 is a pretty value-less card unless you are doing gpgpu or need the small usable ipc difference (~3%) and higher tdp allowance for clock potential similar to the 670/680 (which may net you a another 5%). It has 2048 shaders and the wide bus because when it is clocked to >1100mhz 2048 shaders can use more than 256-bit/7gbps for compute, while also securing it will always be the fastest 32 ROP part per clock. It also doubles as a buffer for getting acceptable yields on a 1792sp part (the most efficient)...unlike gk104 which does that natively, hence we get less-efficient (core) salvage parts and this clockspeed game and nvidia charges you a pretty penny for the parts that all work.

Point being, you have the right to believe what you want and that is fine...but in this case it makes you look ill-informed...or exactly what nvidia thrives on...and I continue to wish people knew better.

Anyway...655 and 650 = probably stock-clocked to equal 7850 and 7770. 650 should be better than 7770, if barely, and 655 is a wildcard only because 7850 is such a funky product (some have low voltage and no voltage control, others bios locked at 1050mhz etc while others have none of those quirks...655 @ 1200 [average for 1.175v] = 7850 @ 1050). I stand by that whatever replaces the 7700 will be clocked to beat these parts; the top model will probably have similar ipc to 655 with higher stock clock, the lesser model more ipc than 650 and a lower clock. I still think ~-900-925mhz and ~1075-1100mhz will be the pro and xt spots going up the lineup for 8000, which would line up well against these cards.

I dont think you understand how to use the TL;DR....

 

[Harlequin_uk]

lol.
TT's reviews are pathetic.
Old drivers, pathetic set of benchamrks - old games, old versions of them
No overclock, no overclocked performance really. No power consumption or noise.
No super high resolutions, no multimonitor setups, and the list goes on and on.
It is funny how the site with the worst reviews gets them first
it makes us, reviewers feel good about holding and releasing the "wow" effect of users responsing to the review and it's results. No worries when it comes to TT.

except that TT have been right the last 4 or 5 times.


carry on fail trolling - it makes you look even more stupid

 

[dj-electric]

I am not the one who troll (wtf trolling has to do with anything here?), i just mentioned a list of facts about the review - aren't they truth?
what is exactly right? right in what? in the games that no one even play today? 4 year old FarCry2?
I want benchmarks in BF3, in Crysis2 and all other new games with new and rellevant drivers. That is not hard.

 

[Harlequin_uk]

your first line ` TT reviews are pathetic` - they are only pathetic if they are badly wrong - and they are not.

 

[dj-electric]

No, they are pathetic for giving information witch a lot of it is irrelevant for most users. And also - not giving a lot of important information as mentioned above.

 

[Harlequin_uk]

so now , because YOU do not like how something is reviewed because of YOUR list , its then pathetic? some people dont care about overclocking , or multi monitors or power consumption - they care about if it works when they plug it in on the 19" screen.

 

[cadaveca]

so now , because YOU do not like how something is reviewed because of YOUR list , its then pathetic? some people dont care about overclocking , or multi monitors or power consumption - they care about if it works when they plug it in on the 19" screen.

ANd some people do care, like he does.

He's entitled to his opinion, and there is no point arguing about it. Might as well just say "I disagree", and leave it at that.

Frankly, I tend to agree with him, but to me it's fine if you enjoy TT's reviews. Many do not, as evidenced by the site's traffic.

 

[Crap Daddy]

some people dont care about overclocking , or multi monitors or power consumption - they care about if it works when they plug it in on the 19" screen.

Those people don't read reviews.

 

[Casecutter]

Their results were later confirmed by other review sites when NDA lifted. I consider it as "helpful". I expect the same with 660 review.

Tweak Town.com that was their "pre-released" 5 day early (un-authorize card), which has a boost clock of 994Mhz, while the released reference is at 980Mhz.

On to this mangle mess... Let's see if I can break-it out better than btarunrs’ irreconcilable way of enlightening us… it can be made rational!
650 on the GK-107 2-SMX 128-bit for $140 that probably won’t beat a 7750
650 Ti on GK106-200 3-SMX (of 5) with 128-bit $160 but can’t reach a 7770
655 with a GK106-250 4-SMX (of 5) 128-bit for what $180 and par with a 7770
660 GK106 5-SMX 192-bit like a 7850 but at $210

Yep, lots and lot's of geldings to "off-load" to the unsuspecting that can’t muddle through the mud.

 

[TheoneandonlyMrK]

Im surprised they arent all called 660 either GT GTX GTI GS, SE, TDI or something, they are actually different, wont that confuse fans

 

[jihadjoe]

No, they are pathetic for giving information witch a lot of it is irrelevant for most users. And also - not giving a lot of important information as mentioned above.

How is the information irrelevant? You won't get exact data on how the reviewed card performs in the most up-to-date titles, but you do get an idea of how it performs relative to other cards that you are familiar with. And those cards you probably know well enough to gauge where the new boy stacks up.

When it's not possible to re-test some of the old hardware (because you returned the card to whoever lent it already), it makes sense to keep data from an old test. TPU uses an even more outdated version of Heaven (2.0) compared to TT (2.5), but nobody complains about it.