7970: Another Disappointment from AMD

[Dangerous_Dave]

It seems like AMD can do no right recently: their Bulldozer chip was designed for a world that doesn't exist, and consequently is sometimes even outperformed by the old Phenom 2 Quads. Now we have this 7970, which by all account seems to be about 30-35% faster than a 6970, and (in terms of minimums) is only just ahead of the 40nm GTX580!

At 28nm you have roughly double the transistors per area compared with 40nm, yet the 7970, which is almost the same size as the 6970, is only 1/3 faster. It should be at least twice as far ahead of the 6970 even on immature drivers. Given that the 40nm GTX-580 is only just behind the 7970 specifically for minimums (which is really the only thing that matters), AMD is surely going to get destroyed when Nvidia moves to 28nm.

Think about it this way: For 3.4 billion transistors, AMD could have done no research at all and simply integrated two 6870s onto a single die (similar to 5870 vs 4870), ramping the clock speed up 20-30% to somewhere over 1Ghz (28nm would easily allow this). This would have produced performance somewhere close to a 6990, and far ahead of the 7970. Instead, AMD have spent lots of money on research and used 4.1 billion transistors to produce performance far worse than a 6990.

There has to be something wrong with a company when the fruits of their research are far worse than if they hadn't bothered doing any.

 

[Dangerous_Dave]

Read other reviews that show a 40-50% gain over the 6970 and 20-30% gain over the GTX 580. In some instances an overclocked 7970 is up there with a 6990, and that's on restricted settings in Overdrive.

Oh and a lot of transistors were spent on Compute performance, not just Gaming performance.

@Compute performance: Why? Couldn't they build separate subtly modified highly priced chips for that purpose?

@20-30% gain over GTX-580: Did those reviews show the figures for minimums? Can you link?

 

[Dangerous_Dave]

What's with all the armchair engineers these days? Not sure whether to laugh or cry after reading this pile of complete and utter nonsense.

The other option would be to offer an intelligent rebuttal. Unless of course you're trying to audition for a job at AMD.

 

[Dangerous_Dave]

Can somebody please link me all these great results for 7970 compared with GTX580? I would hate to think that people were plucking single results out at implausible resolutions/settings and/or not comparing minimums as in original post.

 

[Dangerous_Dave]

It's just the price, wait and see what it sells for before making a judgement.

Regardless of the price, considering the amount of silicon and the technology (28nm vs 40nm) involved, as per my original post, this has to be considered a massive disappointment. Unless somebody can provide links to the contrary.

 

[Dangerous_Dave]

I can't quote every reviewer but I'm seeing a trend.

What I would point out is the when AMD went from the 4870 to the 5870 they also bumped up the TDP from 150 watts to 200 watts plus they had a die shrink to go along with it. The 7970 only has a die shrink and a new design and no extra power to play with it, typically the big increases we have seen over the years have meant a big increase in power consumption so AMD should be congratulated for making a very efficient design that gives a us a big boost in performance and without the extra heat and power.

Most of those quotes don't really give any numbers and are just a lot of hyperbole. Some appear to suggest that in some cases the 7970 only trades blows with the GTX-580. If that's an indication of the best defence that can be given for the 7970, then I'd say my original comments were vindicated.

 

[Dangerous_Dave]

Mmm people aren't really getting that I'm not talking about the selling price at all: selling prices can go up and down depending on all kinds of strategies and business conditions. I'm purely interested in what AMD have been able to achieve with their technology per mm^2 of silicon, which will impact on their profits - what they choose to charge the consumer is another issue entirely. If AMDs silicon is uncompetitive with Nvidias, ultimately they will either have to make a loss or charge too much.

 

[Dangerous_Dave]

There is no intelligent rebuttal to make, especially to a post like yours. Plus, 99.99% of the people on this forum don't know squat (including me) hence it's mostly futile anyway. Crticise the end performance, fine, but I find it absurd when people criticise the actual design process, as if they could have done better, lol. In this respect, as I said before, your comments about transistors, clock speeds, etc is complete nonsense. I don't claim to know better, but I trust AMDs design thought processes far more than yours.

In summary: "your post was nonsense but I don't really know why". It is a terrible argument to say that just because you're not an expert in a field you can't criticise experts. Companies are assessed according to their relative results (aka 'competitiveness'), and if one is falling behind another in a design driven product, it's fair to say that their experts aren't performing.

 

[Dangerous_Dave]

It's just a matter of perspective:

Viewed from a "historically blind" perspective, they are great cards. The fastest GPU around, decent power consumption - what's not to like? If the reviews are taking this perspective then they are absolutely correct. After all, they can't judge the cards against people's expectations, or against the speculated performance of yet-unreleased hardware.

Taking into account the norms of GPU technology though, they haven't lived up to the expectations that many people had of them. The drop from 40nm to 28nm is (proportionally) the biggest we've seen in a long time, but that hasn't translated into performance the way that it has with previous die shrinks.


One thing has struck me though:
The 28nm process allows (in principle) a relative increase in transistor density of 2.05 times (i.e. +105%), whereas the actual transistor density has increased by 'only' 74%. This is quite out of character from previous generations, which have more closely followed the expected transistor packing density (you can check the numbers on the wikipedia pages).

It will be interesting to see how Kepler fares in this regard. It could be that the more intricate design of GCN requires a somewhat looser transistor arrangement than the VLIW cores. But, if nvidia are also producing a ~70-80% increase in transistor density then it's more likely indicative of issues with the 28nm process. That's not necessarily to say it's down to issues with the manufacturing process at TSMC - it could be more physical problems associated with running high-speed transistors at such small sizes. If this IS the case, then we might not see such large gains from Kepler either.

This is an interesting response, and something that seemed odd to me too. As the poster below it indicates, AMD has only increased (useful?) transistor count by 64%, yet we would have expected around 100% by consideration of area alone. This seems to suggest that there is a lot of redundant silicon on the 7970. I read an article a while back that said that in the face of poor manufacturing processes, designers will double or triple up important elements that might not fab right, leading to redundant silicon. Not doing enough of this on an immature process is what led to nvidia's problems with the GTX480/470 release. So it could well be that 28nm is still not working at all well. By the time Nvidia come to market it might be working rather better, which would be bad news for AMD.

Against the most favourable reviews, I suppose an increase of ~40% on 64% more transistors isn't all that bad, especially if they can improve it with better drivers.

So perhaps it's a combination of a slight design fail and a big process fail.

 

[Dangerous_Dave]

Bear in mind that the 6970 will already have had redundancy built into its area usage, so the 7970 must have massive redundancy to fall so far off even the 6970 figure in terms of transistors per mm^2. I guess we'd all been reading about the troubles at TMSC though.