Nvidia say they're underwhelmed by 7970

HighlandeR · 22 Jan 2012 at 23:42

The_Baron said:
Suggests April release rather than the rumoured February. AMD will sell a good few in the meantime

thats all I read lol

Sorta the main thing aint it.... ATI will win almost 1st half of 2012 by the time Nvidia get the full range out then.

Perhaps Aprils ivy launch can be coupled with Nvidias cards

Stripes · 23 Jan 2012 at 04:58

The more strongly quoted rumours I read, the less this whole thing makes sense. A GTX 660 priced at <$300 which would make all >$300 cards incthe 500-series currently on the market obsolete? Kepler doubling performance on a new process without significantly increasing transistor count and cost? Fudzilla being *right*?

Nah, I'm sticking to guessing a late March launch for any Kepler product at the earliest, and being roughly 50% faster than the parts they're replacing.

Conor Royston · 23 Jan 2012 at 05:27

There is also this coming out in March, or apparent March

http://videocardz.com/ati/radeon-7000/radeon-7990

djxput · 23 Jan 2012 at 08:04

tell you the truth I was a bit underwhelmed by the whole 7970 issue and do not find it a worthy upgrade to the radeon line.

I mean didnt people say we are moving 2 steps forward in architecture and meanwhile only barely getting a speed increase? ...

I hope nvda's cards are a true upgrade and something to be desired instead of a so-so upgrade that breaks the bank more then the fps barriers.

lollyhayes · 23 Jan 2012 at 08:13

Where's Drunken Master? This thread should have at least 50 more posts full of his ranting.

xsistor · 23 Jan 2012 at 08:29

lollyhayes said:
Where's Drunken Master? This thread should have at least 50 more posts full of his ranting.

Clearly this thread fails the notability test...

dave_beast · 23 Jan 2012 at 09:29

djxput said:
tell you the truth I was a bit underwhelmed by the whole 7970 issue and do not find it a worthy upgrade to the radeon line.

I mean didnt people say we are moving 2 steps forward in architecture and meanwhile only barely getting a speed increase? ...

I hope nvda's cards are a true upgrade and something to be desired instead of a so-so upgrade that breaks the bank more then the fps barriers.

You clearly know nothing!

CAT-THE-FIFTH · 23 Jan 2012 at 10:07

Stripes said:
The more strongly quoted rumours I read, the less this whole thing makes sense. A GTX 660 priced at <$300 which would make all >$300 cards incthe 500-series currently on the market obsolete? Kepler doubling performance on a new process without significantly increasing transistor count and cost? Fudzilla being *right*?

It seems there are loads of yield issues for TSMC and GF at 28NM:

http://www.techpowerup.com/159189/28-nm-Struggles-TSMC-amp-GlobalFoundries.html

This will probably push up cost and decrease availability too when compared to 40NM.

It makes sense why a load of AMD and Nvidia GPUs are being rebranded this year.

Stripes said:
Nah, I'm sticking to guessing a late March launch for any Kepler product at the earliest, and being roughly 50% faster than the parts they're replacing.

That sounds about right.

I had a look around on techPowerUp. It seems after the 8800GTX, both AMD and Nvidia have improved performance roughly 40% from their refined second edition cards of a generation, to the "raw" first edition cards of a new generation. GTX280 was roughly 40% faster than a 8800GTX. HD4870 was roughly 50% faster than a HD3870. HD5870 1GB was roughly 40% faster than an HD4890. GTX480 was roughly 40% faster than a GTX285. HD7970 was roughly 40% faster than an HD6970.

According to techPowerUp the 8800GTX was a 70% improvement over the 7900GTX.

drunkenmaster · 23 Jan 2012 at 10:44

CAT-THE-FIFTH said:
It seems there are loads of yield issues for TSMC and GF at 28NM:

http://www.techpowerup.com/159189/28-nm-Struggles-TSMC-amp-GlobalFoundries.html

This will probably push up cost and decrease availability too when compared to 40NM.

It makes sense why a load of AMD and Nvidia GPUs are being rebranded this year.

That sounds about right.

I had a look around on techPowerUp. It seems after the 8800GTX, both AMD and Nvidia have improved performance roughly 40% from their refined second edition cards of a generation, to the "raw" first edition cards of a new generation. GTX280 was roughly 40% faster than a 8800GTX. HD4870 was roughly 50% faster than a HD3870. HD5870 1GB was roughly 40% faster than an HD4890. GTX480 was roughly 40% faster than a GTX285. HD7970 was roughly 40% faster than an HD6970.

According to techPowerUp the 8800GTX was a 70% improvement over the 7900GTX.

First part in the story quoted shows what nonsense it is. Some people just can't comprehend the difference between fab output and low yields of a wafer. Nvidia screwed up their design at 40nm nothing more or less, the fact the guy (the ONLY guy in the world) insisting there are low yields claimed Nvidia had a problem on 45nm...... when they made no such product, gives that all the credibility it deserves, zero.

As for the later, your numbers simply aren't correct, Techpowerup for years their round up numbers do not add up with ANY other review ever, their comparison numbers are rubbish, fact, they do not represent reality at all.

The 280gtx was NOT 40% faster than an 8800gtx in any way shape or form.

THe problem with techpower up is, there will be 5 games where the 280gtx is 60-80% or so faster than a 8800gtx and 5 cpu limited, REALLy old games or 3dmarks(they were still doing 3dmark 03 I think it was up till about a YEAR ago), which would have at the time shown next to no difference between a 4870, a 4870x2, a 280gtx, or a 295gtx. Thing is that will average down to bring the supposed comparison performance massively down.

kitch9 · 23 Jan 2012 at 10:50

In other other news, Dog turd stinks, pigs cannot fly and the other day someone actually saw a bear doing a number two in the woods.

Will wonders never cease.

CAT-THE-FIFTH · 23 Jan 2012 at 10:57

xsistor said:
Clearly this thread fails the notability test...

Not any more!

You can "thank" me later!

xsistor · 23 Jan 2012 at 11:04

Yep things are starting to look better around here. We may even hit 5 pages now...

xsistor · 23 Jan 2012 at 11:20

http://www.pcgameshardware.de/aid,8...Aussicht-Geruecht-des-Tages/Grafikkarte/News/

GK110 A2 stepping chips photographed according to this site.

Source also had this

IT was posted on a CHINESE forum and reported by a GERMAN site. So surely it can;t be fake.

Bobisuruncle54 · 23 Jan 2012 at 11:57

6bn transistors?

That's 40% more than the 7970s 4.3bn which has a die size of 365mm2, so we're looking at a pretty large die larger than 500mm2, assuming Nvidia hasn't been able to pull off some density magic (AMD are typically more efficient in this area).

Seems a bit odd though, as doubling the shader cores wouldn't simply double the size of the chip, and comparing the GTX 580's 3bn transistors and ~520mm2 die size, maybe the shaders are more complex or more redundancy has been built in?

If they still have hot clocked shaders then seems as though a 680 could be upwards of 50% faster than a 580, looking at the image it certainly looks like the blurred shader clock is 4 digits long and the GPU clock is 3. But if the rumours of them being dropped are true, it could be more likely it's 30-50%.

Bleh, too many unknowns. Certainly looks as though Nvidia isn't going to shy away from producing big chips anytime soon though.

xsistor · 23 Jan 2012 at 12:01

It's probably fake. But if it isn't, this is the GK110 which is expected to be rather large. It's the GK104 that will be smaller and compete with the 7970.

Bobisuruncle54 · 23 Jan 2012 at 12:42

xsistor said:
It's probably fake. But if it isn't, this is the GK110 which is expected to be rather large. It's the GK104 that will be smaller and compete with the 7970.

Going by Nvidia's past trends, I don't see this being any less than 500mm2 and possibly as much as 550-580mm2.

Has there been any word on how good TSMC's 28nm process is? My problem is that it the 7970 is being sold at a price as if it is the fastest card of the 40nm generation. It just raises a bunch of questions and I wouldn't be surprised if the 7970 is quickly phased out with a card with more shaders and/or higher clocks or dramatically reduced in price when competition arrives.

Edit: Pretty interesting read about the state of 28nm. Seems difficult to tell how it is influencing prices though.

Duff-Man · 23 Jan 2012 at 13:16

xsistor said:
http://www.pcgameshardware.de/aid,8...Aussicht-Geruecht-des-Tages/Grafikkarte/News/

GK110 A2 stepping chips photographed according to this site.

Source also had this

IT was posted on a CHINESE forum and reported by a GERMAN site. So surely it can;t be fake.

Interesting... Numbers are all at least in the right ballpark.

The weight of evidence from recent 'leaks' now seems to suggest that the hot-clocked shaders remain for Kepler. The GK104 was suggested to be 780Mhz core, and now this GPU-z shot seems to suggest shader clock >1GHz and core clock under 1GHz. Who knows though - they could all be fabricated.

Regarding transistor count, die size etc: If Nvidia keep the same packing density as with Fermi, then the downsize to 28nm should have a 6Bn transistor GPU at around 507mm^2. Given the realities of dealing with a new process, I'd suggest more like 530mm^2 is likely for 6.0Bn transistors.

Bobisuruncle54 said:
Seems a bit odd though, as doubling the shader cores wouldn't simply double the size of the chip, and comparing the GTX 580's 3bn transistors and ~520mm2 die size, maybe the shaders are more complex or more redundancy has been built in?

Why not?

Remember, you don't just 'double the number of shader cores' - the entire GPU must be scaled up to prevent one aspect bottlenecking the data flow. if you double the number of shaders then you need, in principle, to double the number of TMUs, ROPs, cache, interconnect logic, etc. If the memory interface width is also increased, then the new GPU will be not too far off double the number of transistors. Usually it will take a few extra transistors to support the increased complexity of the GPU, but an architectural redesign can often negate that.

Take the example of the 4870 -> 5870: Here we saw a doubling of the number of stream processors (800 -> 1600), as well as a die-shrink (55nm->40nm). The number of transistors increased by 2.24x: From 956M to 2150M.

If anything, I would have expected a 1024-shader Kepler card to be at least 6Bn transistors, perhaps closer to 6.5bn.

stulid · 23 Jan 2012 at 13:32

Over at Xtremesystems they say its fake as GPU-Z doesnt have the info for these cards.

Duff-Man · 23 Jan 2012 at 13:32

stulid said:
Over at Xtremesystems they say its fake as GPU-Z doesnt have the info for these cards.

Ah

Bobisuruncle54 · 23 Jan 2012 at 13:47

Duff-Man said:
Why not?

Remember, you don't just 'double the number of shader cores' - the entire GPU must be scaled up to prevent one aspect bottlenecking the data flow. if you double the number of shaders then you need, in principle, to double the number of TMUs, ROPs, cache, interconnect logic, etc. If the memory interface width is also increased, then the new GPU will be not too far off double the number of transistors. Usually it will take a few extra transistors to support the increased complexity of the GPU, but an architectural redesign can often negate that.

Take the example of the 4870 -> 5870: Here we saw a doubling of the number of stream processors (800 -> 1600), as well as a die-shrink (55nm->40nm). The number of transistors increased by 2.24x: From 956M to 2150M.

If anything, I would have expected a 1024-shader Kepler card to be at least 6Bn transistors, perhaps closer to 6.5bn.

Good point, mind is being slow today.

Although it's most likely fake, it seems likely that if we see a doubling of cores we are going to see another huge chip, which makes me even more curious about the 7970s pricing.