AMD RV770 refresh -> RV790

What, you mean deleted? And then the freed die space used for clusters?

Jawed

 

Or replaced by an x86 core with perfect scaling?

(Sorry, rpg, couldn't resist! )

 

If ppl think that sideport done right is sideport deleted, it must be one universally disliked thing.

I meant, some stuff rejiged to make a non afr (or better sfr) x2 card, even if they dont launch a 790x2.

 

Will explain more in that thread. Clearly what you have in mind i very different from what I am thinking. A detailed explanation of my pov should help.

 

With the current economic environment, if you don't offer real value for price (who in the world would buy a card which only has a better idle consumption for, let's say, 50$ or € more?) you will experience serious problems in selling your products...

Then, again, with R420 there were yield issues, so the transition to R480 was an advantage for the company, but now, what kind of advantage will AMD get from such a chip (given that it doesn't seem that RV770 has a problem with yields...)?

Develop even a chip revision could be very expensive, so why take the risk just to improve consumption or only to slightly rise clocks? If I had to finance something like that, I would invest to further move the market (i.e. targeting higher positioned VGAs...).

 

To not leave the (rather silly) performance crown to the GTX295 probably? While you do have quite a few legit points I'm rather curious at this stage what exactly is going on with TSMC's 40nm and if it's safe to use it for chips with higher complexity at the moment.

I don't think RV790 marks just a frequency increase over RV770 (the estimated performance increases don't make sense for that), but on the other hand a roughly 30% increase (via slightly more units + added frequency) isn't outside the ballpark of your above reasoning.

It makes sense to fill the "grey area" between RV770 with something until their D3D11 generation.

 

I guess that depends. For example on whether or not or for how long you will/can allow those two products to coexist in the market. If you can limit this time to a pretty short amount, then it would really be cool to improve upon the margins a bit.

BTW, please note, that I said, that the addition of a better idle mode would make the product more compelling than just raising clocks, so i didn't exclude that from happening.

I am not familiar with the delicacies of all of TMSCs process nodes at 55nm listed here, but I'd assume "Vt" standing for voltage, an the GC-option offering a different choice including something called "ultra high"…

Again, I am not sure how costly it would be to just port a given design from 55GP to 55GC. Maybe, if it's not too expensive, AMD feels that it is the right thing to do to bridge the time to coming 40nm-GPUs.


--
I just want to make clear again, that I am not thinking predisposedly in any direction right now. I do not consider it a given that AMD may or may not stay with the same chip nor that they necessarily had to add additional SIMDs/RBEs to improve in performance in the rumored percentage range.

 

Vt is for the threshold voltage. The useful current through a CMOS transistor ~ (Vgs-Vt). In most situations Vgs = supply voltage. So the higher you make Vt, the lower your performance. But the higher your Vt the lower the leakage...

There's no free lunch...

 

And there's an "Ultra High" option for that in 55GC. Damn, now what to make of this? If it's opted for at all, my money would be on extra SIMDs but lower clocks and superior idling.

Fun!

 

If you're happy with max clocks of 200MHz: sure!

 

As silent_guy said... Plus, I think it's worth repeating again and again that low-power processes result in higher power in active mode for a given level of performance (higher voltages for a given frequency...)

 

Yes it's new for this chip. Designed for higher performance. Costs more than 55GP. Look back at 90nm had GP, LP, GC and GT versions i think.

Is some risk for amd to do as costs are higher relative to RV770(process cost and slightly larger die) and there likely will be a high degree of substitution. ie end up selling a RV790 instead of RV770 rather than RV790 over nothing or RV790 over GT200. The RV790 better be able to sell for significantly more than RV770 or will not make back development costs.

 

Well in the first place i care about performance which i can use. If they can achieve this in the kHz range - i'd be happy with that either. Point taken - I didn't realise that it'd be that limiting.

 

Well, don't pin me down on that number! I had never heard of a Very High Vt process before, but low power processes carry a significant speed penalty, so this one sounds especially ominous.

 

Of course not, but I was thinking more in terms of maybe 10-15 percent lower clockrates.

 

Expreview writes HD 4890 will clock at 950/900 MHz.

 

Makes slightly more sense in terms of performance. Yes, GTX280 league.


What about leakage?

 

A few sites have been saying 55GT process...
Leakage hasn't really been a huge concern with the current 55nm process they are using.

 

RV790 seems to come with a known quantity of ALU-lanes: 800, albeit mightily pimped clockrates of 950/900, so sez expreview.

http://en.expreview.com/2009/02/23/...-more-than-super-super-overclocked-rv770.html

edit:
Oh sorry, a bit late.

 

Even so, 950 on a stock chip is definitely uncharted territory.

The 3870X2 only went as far as 825 IIRC, but the same process that the RV670 (from what I heard from sauces 55GT was used then) has been built on should have been refined since then.