Howdy, Stranger!
It looks like you're new here. If you want to get involved, click one of these buttons!
Quick Links
Radeon HD 9000 series rumors: still on 28 nm process node means not much of an improvement
There are rumors that AMD will launch their Radeon HD 9000 series cards in about 2-3 months. Yes, 9000, not 8000; the 8000 series is OEM-only rebrands of 7000 series parts. 20 nm isn't ready yet and seems to be coming slowly, which means that the cards will still be on 28 nm. The existing Bonaire card, the Radeon HD 7790, would likely get rebranded into the 9000 series.
I'd expect the lineup to include the old 7790, a cut down version of it, and a more efficient big GPU chip that can compete with the 7900 series (since the Tahiti chip is a power hog). AMD may or may not have a high end chip to compete with Nvidia's GeForce GTX Titan and GTX 780. There's also a decent chance that AMD will launch a new low end chip to replace the old Radeon HD 6670; Sapphire is rumored to launch a 384-shader Radeon HD 7730 soon.
Really, though, this is bad news, as it looks like the 20 nm cards that would be the real advance aren't coming anytime soon. You don't launch new 28 nm cards and then move to 20 nm three months later; if you were going to do that, you'd skip the new 28 nm cards. All of the foundries seem to be struggling with 20 nm, though TSMC is promising a 16 nm FinFET process node only a year after 20 nm, while Global Foundries says that they'll have their 14 nm XM node only a year after 20 nm. Of course, those dates could easily get pushed back as they get nearer.
Years ago, everyone seemed to agree that getting to 22 nm or so was doable, but the question was about beyond that. While it's still coming, the "beyond that" seems to be troublesome for everyone. Or at least for logic chips; there is already 19 and 20 nm NAND flash widely used in many products.
Comments
That's not necessarily true.
If you just look at the top end card:
5870, 6970, and the 7970 (remember they adjusted the naming convention there)
Those 3 cards are all radically different, with very different architectures. The first 2 were 40nm, and just because they didn't move the process node, the 5870 was a nice card, the 6970 was still a marked improvement because of additional technology that was added (PowerTune, among others).
You don't need a process node jump to make a generational push meaningful or relevent. They also don't need a card to compete with Titan, at least until Titan-level technology is down in the <$500 price bracket.
Rebranding is a marketing tool.
If your competition gets ahead of you, and your same old cards have been sitting on the shelf for 12-18 months, then you can get a small burst of wind just by updating the clock speeds and changing the name.
Its sleazy, but a lot of people fall for it, and it's far from the worst evil video card marketers have done.
I really didn't think they needed to rename the 6000 series cards as well. It was marginal boost between generations where the 6870 performed worse than the 5870. Between the 3870, 4870, and 5870 you could see significant improvement with each generation. Having the 7970, 8970, and 9970 being exactly the same card makes no sense to me at all.
The thing I really take from this is the 9970 was a card developed for a 20nm process node. Remaking the card for a 28nm process node simply won't work. You will end up with a gigantic and hot chip.
AMD are keeping the specifications of this card so secret it will make a Bilderberg meeting look like a free public event.
No one knows, the only sure thing to bet on is that it's going to a true DX11.2 GPU.
I'd expect some slight efficiency and architectural improvements, but only slight. Remember that we've already seen one of the new generation of cards in the Radeon HD 7790. If you can imagine a card with the performance of a Radeon HD 7970 together with the energy efficiency (performance per watt, not total power consumption) of a Radeon HD 7790, we'll probably see something like that.
While Ridelynn is correct in the abstract that a new architecture can be important even on an old process node, this isn't going to be a hugely important new architecture. Probably the most important architecture change on an old process node that comes to mind is that Conroe (first Core 2 Duo) was on the same process node as Cedar Mill (last Pentium 4). But it's easier to make a huge improvement over your old architecture if the old one was terrible, which NetBurst (Pentium 4 and Pentium D) was and GCN (Radeon HD 7000 series) is not.
Well these are supposed to be GCN 2.0 - whatever that may bring to the table I'm not entirely certain. I would bet a better PowerTune - more inline with nVidia Boost 2.0's more aggressive temperature compensation. They already have Zero-Core for completely idle GPU's, I'd expect something additional for low power state power draw.
I've seen rumors of a ~20% increase in per-core performance going from GCN 1.0 to 2.0 with similar TDPs, which wouldn't be out of the ballpark from just a efficiency refinement standpoint.
Rumors contradict a bit though - some say the 8000 series will be GCN 2.0 based but stay with 28nm, some say they are just rebrands at GCN 1.0 and the 9000 series will be GCN 2.0 based (at 28nm), and others say the 9000 series will be the first GCN 2.0 products at 20nm.
Q3 isn't that far away though, and that is one piece of information that is consistent in all the rumors.
Compare a 7970 to the newer 7790 and we've already got those improvements. Getting the same improvements in the rest of the lineup doesn't seem like much of a stretch, but neither is it a huge deal. Still, the 7870 and 7770 are already more efficient than a 7970, so efficiency improvements as compared to Cape Verde (7700 series) or Pitcairn (7800 series) cards will be less than as compared to Tahiti (7900 series) cards.
You'd expect efficiency improvements on the order of 40% from a full node die shrink, and we're not getting that. It does make me wonder when we'll see 20 nm cards, though.
The 7790 AMD still calls "based on GCN", but it has evolved some from the original GCN 1.0 clearly - but not so much that they have come out and called it 2.0. Will it be the final GCN we see in the next generation? I don't think so, I think we'll see a bit more evolution based on what we see now in the 7790.
http://www.anandtech.com/show/6837/amd-radeon-7790-review-feat-sapphire-the-first-desktop-sea-islands/2
And this rumor, which if it's true, has AMD calling something GCN 2.0
http://forums.anandtech.com/showthread.php?t=2230338
It's very common for a product generation to be using multiple various architectures.
In fact, I can't think of many generations where there was a homogeneous set of architectures across every product in the lineup.
All that being said - yes, a very large portion of it is smoke and mirrors.
If the ~typical~ consumer goes out and see the
nVidia 610 versus the 480 - which one are they going to pick? Probably the 610, it has a higher numbers - and that's no joke, I see it all the time.
AMD is just as guilty - the entire naming convention shift between the 5000 and 6000 series didn't help at all either, because just as mentioned before, it even confused people who were otherwise literate and knew the naming convention. A 6870 was slower than 5870, and the 5970 was dual GPU card that was faster than the 6970, but only because it was dual GPU - it really made things confusing.
And again, if the ~typical~ consumer is out to buy a card, the 7750 looks a lot more appealing than the 5870. It's a bigger number, it has more buzzwords, and likely whoever makes the card will put a CGI-enhanced picture of a woman with big guns (in both senses of the word) on the cover.
And then there's Intel, who can't even get enough differentiation between their graphics solutions to really make a difference.
Also, with regard to the 7790:
It's also common for GPU makers to release an interim product in between generations. nVidia's 460 Fermi-fix is a perfect example; where just about every other Fermi card in the 400 series was a disaster, except the 460 which released a few months later (and only a few months before the 500 series respun Fermi into something usable). The 460 comes out, fixes a big gaping hole in the product lineup, and previews a lot of the new technology we then saw with the 500 series Fermis.
We see a similar thing with the 7790 now - The GCN 7970 isn't nearly as far behind the curve as nVidia's 480 Fermi was in the first place, but it's still a similar situation. AMD has some fixes ready to go - they tightened up a hole in their product lineup, and Quiz is right - we probably will see everything the 7790 brings to the table scaled up into the next product line (and maybe some more on top of that).