AMD: R7xx Speculation

But RV740 could be created by deactivating parts of RV770. And RV730 looks like a redundant chip, since its specs are way too close to RV670.

 

RV670 on 45nm.

 

wow! Imagine how hard for people is to grasp possibility of 800SP's when this info that’s THREE months was easily forgotten and buried under this gigantic thread!

 

I hate to bring up an obvious question, especially since I do like (and I've already mentioned earlier in this thread) AMD/ATi's tighter control over it's internal's.

However, even assuming the slide is from AMD, do we have any reason to believe this depicts the real specs of the RV770 & isn't another attempt by AMD to try to flush out leaks? Since leaks happen more often the closer to a product release, I could easily see them setting up a scenario like this.

 

Another performance numbers :

 

Sure its possible... but I say that what likely happened was they had press event to unveil the cards

Think about when Nvidia held their press day event and soon we started getting leaked slides from CJ and others

Well, who's to say ATI didn't hold theirs recently? After all, at first TGDaily said that the card was delayed to June 23rd then all of a sudden people were confirming that it was to be released June 25th, a different day, and all at once...

 

Didn't someone under NDA (w0mbat?) say that the launch was in fact never delayed and that the release was planned for June 25th from the very beginning?

 

Yes which is why he (and others) dismissed those fake documents w/ 480 SP's and split shader cores and so on immediately and early but fell on deaf years

And besides, if AMD was going to fake numbers, why would they put numbers that make their card look better than it actually is? It makes sense to underrate your card when faking #'s, not overrate it then disappoint everyone...

 

Not really since the thread scheduler is supposed to pack instructions into each 5 SP cluster. So if for example you had one that was 3 and one that was 2, it would fit into one cluster.

Or it could (in theory) pack say 10x3 into 6 clusters. So in theory there would be no loss in going to 10 SPs per ALU. That same 10x3 would then fit into 3 clusters. Although I highly doubt this would be introduced as I'd imagine it would increase the scheduling/packing complexity.

Regards,
SB

 

I think 4870 was always planned for the 25th (well, once ATI finally set a date). What's in question is whether ATI had set a date for the 18th for 4850.

Personally, I think ATI was still considering a launch on the 18th due to a couple things. 1 - They apparently have stock of 4850 already and could reap an extra weeks sales. 2 - Pressure from AIBs to launch earlier than 4870 since they have stock and need the money.

Going against a launch on the 18th, would be the fact that you're launching your mainstream part a day after Nvidia launches their enthusiast parts. That's a losing proposition right there as it'll be buried under all the hoopla of the Nvidia launch.

Launching ATI's mainstream and performance-mainstream parts together, a week later than the NV launch at least gives some time for the excitement of the NV launch to die down a bit. And if supply is an issue for 260/280, it makes the ATI launch look even better. As well, comparing 4870 to 260 looks a lot better than comparing 4850 to 260.

All in all, a launch on the 25th makes a LOT more sense for ATI once they could no longer launch 4850 ahead of 260/280.

Regards,
SB

 

Exactly... you know your cards aren't meant to compete in the same bracket, so why give the illusion you are and that you failed again?

 

I have to agree with the 800ALUs and 40TMUs.
And about the 80 10way units...

 

Deactivating 1/4 of the ALUs and TUs is brutal - seems extremely unlikely they'd routinely go this far. I'd expect deactivation to be on a smaller scale.

Though I will admit I've got no idea how deactivation is applied in R6xx GPUs. The only thing I know about is deactivating memory channels.

If a TU needs to be deactivated then it has a knock-on effect on all the SIMDs. That's a very serious cost, so that's why I wonder if AMD would even bother with that kind of redundancy.

If it's equipped with a 128 bit memory bus and since it only has 2 RBEs, it should be a fair bit smaller than RV670. Let's say 150mm2 for argument's sake...

I revised RV710 to be 2x RV620. RV730 versus RV635 is:

• 2.67x ALUs
• 2x TUs
• 2x RBEs

I'm wary that the 2-way splitting (green and orange sets) is a costly overhead, but it's attractive in terms of batch sizes...

I'm also wary that fitting these 4 GPUs under $300 could be quite hard.

Jawed

 

Well, with nVidia chips this seems to be quite common (G80 GTX vs. GTS, GT200 280 vs. 260). But you're right that RV770 is a much smaller chip and yields will probably be good enough, so there will be no need to deactivate its parts.

 

What do you think about this configuration of RV770? Possible?



to compare with RV670:
http://pics.computerbase.de/1/9/6/4/3/23.png


It think this could be a good way between transistor-count and efficiency: only 4 ALU-SIMDs, 5 of the fat TUs of RV6xx but with doubled filter-power.

 

Extremely unlikely, I hope ATi won't revert to vector ALUs

 

So I've been watching this thread for a while, and in light of what I've read so far, just to state the obvious:

25% more die space on the same process... 320sp + 25% = 400sp. 400sp x2 for R700 = the magic 800sp number?

Seems logical, assuming any of these rumors are actually true.

 

Ninjaprime: Could you explain, how R580 could carry 3-times more ALUs than R520, being only 18% larger?

 

I want double performance in a few shader-bound games. Apparently, it is not plausible.

 

Your reasoning would be correct if the 320sp took 100% of the die area, which is obviously not the case...