You are skewing numbers and not taking into consideration many factors when comparing architectures/products.
Printable View
What I'd expect to see, and being realistic about those expectations, are the 9950 beating / matching the 770 at $399 range and the 9970 in the 499-549 range and slightly behind the 780 (10% +/-). Pending game bundles and aftermarket options they could both end up being better value than Nvidia's offerings.
As you said tajoh, the head room on 28nm has nearly run its course so unless they've made some impressive gains efficiency wise, improvements will likely be akin to the 5870-6970.
I am being realistic here. The only time AMD has exceeded my expectations was the 4870(the 9700 pro days is too long ago) and that was a case of where AMD got to 2.5x everything about the 3870 because the 3870 was so small to begin with. The only time Nvidia has met my expectations were actually this generation and the gtx 8800. Both cases where we got a card faster than the dual chip predecessor and then some.
The gtx 760 and 7870 are very similar component for component. The 7870 has 12 percent more shaders, the gtx has 25 percent more bandwidth and they have the same clocks, but the gtx 760 is 20-25% faster. This is exactly why gk104 is able to catch up with tahiti which has way more shaders and computational power.
Component wise, a 2304 shader caracao xt(the 4024 shader rumor is ridiculous this generation) is going to have a hard time catching up to a 2304 shader gtx 780. GCN 2 better makes some pretty drastic improvements in efficiency all around, without cutting out the computing power.
AMD gave up something with the 7xxx series to make the chip more compute friendly. Unless they are going to cut some of that out, they are going to have a hard time beating gk110. gk110 552mm2 seems well utilized this generation and seems a bit more compute gaming orientated than the last generation. Combining both of these characteristics makes its so AMD has a hard time catching up to them with a 400nm or less chip, even 450nm might not do it. A fully spec'd gk110 has pretty scary performance potential.
If they were willing to push themselves to the point where some efficiency is lost like the 7970 ghz (which is partially made up with turbo) and used the same ddr5 in the gtx 770, they could probably make a card 50 percent faster than a ghz edition(15% faster than what they have now; a extra 100 mhz(some overclocked cards already have this like the evga acx), 14% more bandwidth and 192 more shaders). The gtx 580 or gtx 285 had no where near this potential.
A HD 9870 (pitcairn update) with 1792 sp , 256 bit memory at 7 Ghz (224 Gb/s) , 32 rops and improved front end (4 ACE, 3 geometry engines) and clocked at 1.1 - 1.15 ghz will easily match or exceed a stock gtx 770. power consumption will be around GTX 680 (200w TDP). Pitcairn has already shown that GCN (stripped of DP fat) is very efficient. a Pitcairn scaled to compete with GK104 Kepler should be a very close fight in perf and perf/watt.
http://www.guru3d.com/news_story/amd...n_october.html
firstly the current HD 7970 / HD 7970 Ghz chips use the same silicon except for higher voltage and BIOS changes. it was the first GPU chip on TSMC 28nm process and released on Jan 9, 2012. the Tahiti chip must have taped out in late Q1 2011. The design must have been extremely conservative as it was done on a immature 28nm TSMC process. the TSMC 28nm process has improved and so has AMD's learning of the 28nm process. AMD's HD 7790 Bonaire (160 sq mm) was able to fit in 30% more perf at the same 85W TDP of HD 7770 Cape Verde (123 sq mm). for a mere 37 sq mm the front end was doubled and the sp count went from 640 to 896. the power management was improved and the HD 7790 was a very power efficient chip. the HD 9970 / HD 9950 and HD 9870 / HD 9850 chips will incorporate the improved power magament of Bonaire (plus maybe even newer stuff) and have aggressive voltage binning (improved tsmc process) to keep TDP well under control to reach their respective perf and TDP targets.Quote:
Component wise, a 2304 shader caracao xt(the 4024 shader rumor is ridiculous this generation) is going to have a hard time catching up to a 2304 shader gtx 780. GCN 2 better makes some pretty drastic improvements in efficiency all around, without cutting out the computing power.
AMD gave up something with the 7xxx series to make the chip more compute friendly. Unless they are going to cut some of that out, they are going to have a hard time beating gk110. gk110 552mm2 seems well utilized this generation and seems a bit more compute gaming orientated than the last generation. Combining both of these characteristics makes its so AMD has a hard time catching up to them with a 400nm or less chip, even 450nm might not do it. A fully spec'd gk110 has pretty scary performance potential.
If they were willing to push themselves to the point where some efficiency is lost like the 7970 ghz (which is partially made up with turbo) and used the same ddr5 in the gtx 770, they could probably make a card 50 percent faster than a ghz edition(15% faster than what they have now; a extra 100 mhz, 14% more bandwidth and 192 more shaders). The gtx 580 or gtx 285 had no where near this potential.
the HD 9970 Hawaii XT is expected to a 2560 sp chip with the hawaii pro a 2304 sp chip. Tahiti was also a chip with bottlenecks. its front end was not able to keep the sp well fed. the ROPs too were bottlenecked. by improving the front end to 4 ACE , 3 geometry engines and increasing ROPs to 48 the per sp performance can easily increase by 10%. add to it 25% more sp and its not difficult to see a 30% perf increase at the same 250w tdp. remember the HD 7970 Ghz had a stock voltage of 1.25v as compared to HD 7970 which ran at 1.175v.
The Hawaii XT can easily run at 1 Ghz at 1.175v - 1.2v and provide 25 - 30% higher perf than HD 7970 Ghz and stay within 250W TDP with the improved power management and chip binning. if a 560 sq mm , 7 billion transistor, GTX Titan running at upto 1 ghz boost can stay within 250w tdp there is no reason why the hawaii chip which should be around 5.5 billion transistors and 440 sq mm cannot stay within 250w tdp while running at 1 ghz boost clocks.
http://www.anandtech.com/show/6774/n...nce-unveiled/2
"On our sample the top bin is 1.1625v, which corresponds to a 992MHz core clock. Overvolting Titan to 1.2 means unlocking two more bins: 1006MHz @ 1.175v, and 1019MHz @ 1.2v. Or put another way, overvolting on Titan involves unlocking only another 27MHz in performance."
both AMD / NVIDIA naming number make me sick
I think this is more realistic than some of what has been posted in here. But if they are able to compete (or come within 10%) of the Titan, then this will be fantastic news. I was just about to pull the trigger on a couple of 780's. But I think I will hold off to see what sort of numbers these cards can do.
"A few days ago, AMD released its latest Catalyst beta graphics driver ? and hidden within, AMD has helpfully provided a long list of new GPU and APU code names that the driver is compatible with, and provided a little more background on the highly anticipated Kaveri desktop APU. The successor to the Kabini and Temash mobile APUs will be called Mullins. On the desktop, discrete GPU front, we now know that there will be an 8000-series (Sea Islands) GPU called Tonga, and that the 9000-series (Volcanic Islands) will feature Hawaii and Vesuvius GPUs."
http://www.extremetech.com/tag/vesuvius
http://www.extremetech.com/computing...ealed-analyzed
I'm starting to get confused.
So this cards still aren't the real deal? They are the same as those new 8xxx cards, that apparently are only for OEM's, still on 28nm and still just a stopgap.
Are these cards even next-gen-crammed-on-28nm-while-waiting-for-20nm, or just tweaked-old-gen-with-bigger-die? I would say business wise, second option would make more sense with 28nm being more mature and all.
Oh Ok, just read the #16 post on this thread, it's old gen, makes sense.
This is a really interesting approach to a new process that completely throws the "test low end cards first" out of the window.
You first deliver you high end arch, that despite being on a non mature process will still be a competitive decent card, you then start releasing the mid and low end cards, testing new features on them, and also and more important, measuring the process advances, being able that way to better calculate the performance\target TDP of a refresh high end card on a mature process.
Wait what?
OK so I'm confused now. 9000 is coming in October? Did I blink and miss 8000?
I'm looking around for a new card ATM, and was considering 7970, but someone suggested waiting a few months for the 8000 cards. Now I'm reading 9000's are coming soon?
8000 series are rebadged 7000 series for OEMs. So they are skipping and going to 9000 to avoid "confusion"
At a recent sponsored LAN event I attended, power color representatives were asked about the new cards being 20nm, which they quickly confirmed. I cannot verify the accuracy of their answer so take it with a grain of salt.
Interesting. If this is true it's quite the feat.
If true, how did they manage this? Mass production hasn't started yet, so they must have opted for risk production which should be quite expensive.
Actually, thinking about it again I can see how 20nm kind of makes sense.
Originally the product line was supposed to be "stable through 2013" and somehow changed to a major refresh in the second half of the year. If VI is a 20nm product, then it's possible that the first test batch out of the oven came out far better than anyone at AMD expected and they pushed the launch up. It's still borderline unbelievable, but explains the massive shift in AMD's timetable better than a 28nm VI ever could.
If true, given the cost of the new node and the lack of capacity this early on I'd expect a 20nm VI to be very expensive, very supply constrained, or both. Could end up being the mother of all paper launches.
It's so confusing - some people say 20nm, others say 28nm. I hope concrete information is leaked, soon.
20 nano...cue the Jaws theme :cool:
Just want to point out that risk production doesn't necessarily mean that it is more expensive, they would be getting wafers for the same price as mass production, it would just mean that they are taking a slightly larger risk than usual that yields will be acceptable on the process with their design. This would mean either they have a near perfect design for the process or they have built in more redundancy to negate the effects of the new process.
Still finding it pretty hard to believe that the 20nm rumor is getting so much traction. I still find it highly unlikely but not impossible.
FYI- There was always a new product launch planned for 2H '13... Some people just didn't want to listen, looking at you Sky.
Or they could pull an Nvidia and have a few CUs disabled, enable them later on when yields improve and launch a refresh.
Well, thanks to that Power Color rep there's actually some evidence for 20nm now besides just wishful thinking. Still seems incredible though.Quote:
Originally Posted by LordEC911
Can you elaborate? The entire 'stable through 2013' thing seemed pretty cut and dry to me.
The "stable through 2013" slide only included Q3, not Q4.
As for the Power Color reps...who says they weren't lying or just misinformed? Seems unlikely to me that they would give away such an important piece of information just like that.
I thought the other rumor was 22nm. I thought 28nm was all but dead
Although a move like this could explain why kavari has also been delayed a bit.
The idea they are taking a move right out of nvidia's play book by disabling parts of the chip until yields are better makes perfect sense though. It would allow them to operate on a new fab, cores that don't make the cut for 9970 could be used on lower product lines based on how much of the core is good, and it leaves room for a product refresh on the same core once fab yields are better by enabling more of the core later. If you think about it in this sense, the risk is alot lower.
Kaveri is made at GF afaik, so this has nothing to do with TSMC's 20nm process.
1. True, but that would add more power and potential leakage to their ASIC, not to mention the die size. IMO AMD can't really afford to do that.
2. Keep in mind that false/bad info is given out on purpose to diff members of the press and AIBs to try to find leaks.
3. The fact that they were in the process of taping out or had already taped out at least one, probably two, ASICs right around the time of the interview...
Most likely, they didn't know and just agreed. If they actually had heard something, its hard to dismiss the fact that it was probably misinformation.
28nm or 20nm. 22nm was never even discussed/rumored since TSMC skipped 22nm.
Kaveri was rumored to have been pushed back, most likely it was a conservative move for "mass availability."