So, do we know anything about RV670 yet?

[trinibwoy]

For $199, yes. More than enough upgrade from the 8600GTS and sits nicely below the 8800GTS, hence 8800GT. But that plan went haywire as soon as we got reports that RV670 was up and running ahead of schedule. Now their line up is completely screwed up, nomenclature-wise.

It is highly unlikely that Nvidia waited a year to fill the gap in their lineup with a chip with half its units disabled. Think about it for a sec. Why would they plan to replace a 96/128 shader config (8800GTS-320) with a 64/128 one? That make sense to you from a cost perspective?

Btw did you think that Nvidia would launch their refresh minus the flagship?

Depends on whether you consider the GT a refresh or just late to the party.

 

[turtle]

Depends on whether you consider the GT a refresh or just late to the party.

Or both a high-end refresh for now and the mainstream part for when the big daddy launches in Q108 (g100?).

In that case, yes, I do think they launched it without their flagship, and G100 will be alone generation-wise for quite a bit, just like G80 was (6 months iirc)...The difference being G92 is no 7900gt/x, neither in price nor performance. In this case it will work, and work well.

 

[Arty]

It is highly unlikely that Nvidia waited a year to fill the gap in their lineup with a chip with half its units disabled.

They waited because they didnt needed to replace anything, competition drives them and in this case they knew AMD had no answer to the 8600GTS let alone anything between that and the 640MB GTS, which they covered to an extent with the 320MB model. Also note that Nvidia is launching their refresh almost a year after their previous one, when did that happen last time .. huh?

Think about it for a sec. Why would they plan to replace a 96/128 shader config (8800GTS-320) with a 64/128 one? That make sense to you from a cost perspective?.

Makes perfect sense.
64/96 - $199
128 -$299
GX2 (192 or 256!) - $Gazillion

IMO, I believe that 8800GT was originally intended as 96 SP part, effectively replacing the 320/640MB GTS for $100 less. The 128 SP would replace the GTX/Ultra and the GX2 would become the flagship.

Depends on whether you consider the GT a refresh or just late to the party.

GT is G92, Nvidia's refresh to the G80. Its their priority/strategic move that they decided against showing their hand (flagship) early.

 

[trinibwoy]

They waited because they didnt needed to replace anything, competition drives them and in this case they knew AMD had no answer to the 8600GTS let alone anything between that and the 640MB GTS, which they covered to an extent with the 320MB model. Also note that Nvidia is launching their refresh almost a year after their previous one, when did that happen last time .. huh?

You didn't answer my question. Do you actually think that they planned a 64/128 config from the start. What would be the reasoning behind such a part? Expectations of ridiculously poor yields on 65nm?

64/96 - $199
128 -$299
GX2 (192 or 256!) - $Gazillion

Makes perfect sense.

No it doesn't. Look at it this way. What has Nvidia lost by releasing a 112/128 part at $199 instead of your 64/128? Why would they ever plan to release the latter? When have we ever seen an SKU with half the chip disabled?

IMO, I believe that 8800GT was originally intended as 96 SP part, effectively replacing the 320/640MB GTS for $100 less. The 128 SP would replace the GTX/Ultra and the GX2 would become the flagship.

Even if that were the case what has AMD done to motivate Nvidia to bump that SKU up to 112 SP's? RV670XT doesn't look to be anything impressive so far, even compared to the existing 96 shader GTS parts.

We are hearing that Nvidia changed schedules and configurations at the last minute yet nobody has presented any reason why they would do so. Performance of RV670 doesn't seem to be it.......

 

[Geo]

I seem to recall various ATI bargain-bin EOL GTO's shipped with 1/2 the chip disabled. I'd agree I can't think of any examples at the beginning of a chip's existance. But configurability seems to be getting to be a bigger and bigger factor, doesn't it? The chip size of G92 would seem economically untenable by historical standards at the price points NV claim to expect for 8800GT without it.

 

[Lux_]

The thing I don't yet understand: why was NVidia supposedly afraid of RV670 and released mega-8800GT, when RV670 was only supposed to bring "R600 performance with better price": with 256-bit bus and now rumoured 775MHz clock it really can't be much (if any) faster than 2900XT...

It would be interesting to know, what NVidia knows and we don't.

 

[Geo]

It's still a "V" part, so I don't know why anyone would have expected it to surpass R600 performance in the first place (with the possible exception of AA performance if they addressed that).

 

[Bouncing Zabaglione Bros.]

It's still a "V" part, so I don't know why anyone would have expected it to surpass R600 performance in the first place (with the possible exception of AA performance if they addressed that).

Historically, we'd expect an improved refresh at the high end (eg the rumoured R680) rather than a refresh of the value part. This is especially true given the poor reception/yields of R600 and its EOL already happening.

Right now it looks like ATI will have a good performing value part, but not a high end part, which is particularly weird. Maybe AMD decided the OEM part was more important to get out the door on time?

 

[Arty]

You didn't answer my question. Do you actually think that they planned a 64/128 config from the start. What would be the reasoning behind such a part? Expectations of ridiculously poor yields on 65nm?

They did, there couldnt possible be any other reason for naming it GT. It was <=96SP, with possibly the 64SP version taking up 8600GTS's place. So I believe they wanted to refresh with a 96/128 config from the start, with GT replacing GTS and GTS replacing GTX. The GX2 would at a later point replace the Ultra.

No it doesn't. Look at it this way. What has Nvidia lost by releasing a 112/128 part at $199 instead of your 64/128? Why would they ever plan to release the latter? When have we ever seen an SKU with half the chip disabled?

They have effectively messed up their current line up with their refresh.

Because if they did that, they would not have the head ache of bumping up the GTS.

The last question is answered by Geo above.

Even if that were the case what has AMD done to motivate Nvidia to bump that SKU up to 112 SP's? RV670XT doesn't look to be anything impressive so far, even compared to the existing 96 shader GTS parts.

We are hearing that Nvidia changed schedules and configurations at the last minute yet nobody has presented any reason why they would do so. Performance of RV670 doesn't seem to be it.......

RV670 as the naming implies traditionally would point to an "value" sku. However all that changed when we heard that it was up and running ahead of schedule and it was effectively an improved (TBD) R600 and DX10.1. Now, 2900XT is very competitive with the current GTS which is a 96SP sku. Common sense would suggest that if they needed to combat an improved R600, they would need more SP than that .. Makes perfect sense.

 

[ShaidarHaran]

Historically, we'd expect an improved refresh at the high end (eg the rumoured R680) rather than a refresh of the value part. This is especially true given the poor reception/yields of R600 and its EOL already happening.

Right now it looks like ATI will have a good performing value part, but not a high end part, which is particularly weird. Maybe AMD decided the OEM part was more important to get out the door on time?

I don't believe it can truly be said that RV670 is a refresh of anything. AMD has a huge gap in their lineup from 2600 XT up to 2900 XT, only recently somewhat filled with the 2900 Pro. RV670 must have been planned to fulfill a specific market demand, in a particular segment (performance/entry-level enthusiast).

 

[nicolasb]

NV isn't the only one successful at misdirection... Sudden NV roadmap changes at the end of september concerning the final config and price of G92GT kinda suggest that they expected something different (less powerful) from RV670 and most of all they expected it in Q1 08.

That wasn't misdirection: even ATI expected RV670 wouldn't be out till January. They had assumed that the production silicon would be A12, as it is with most ATI parts; they were as surprised as anybody when the A11 silicon turned out to be completely bug-free.

 

[silent_guy]

It was <=96SP, with possibly the 64SP version taking up 8600GTS's place. So I believe they wanted to refresh with a 96/128 config from the start, with GT replacing GTS and GTS replacing GTX.

You're overlooking one really inconvenient detail: statistics.

If you have a chip with 8 identical blocks for redudancy, your yield goes up incredibly fast when you start disabling units and gets very close to the yield that is defined by the blocks for which there is no redudancy.

The number of additional dies you can expect to recover when going from 6 out of 8 to 4 out of 8 will be very low. We know that the yield for a 6 out of 8 die is around 70% (see Arun's G80 review). Let's say it's the same for G92. Even if 100% of G92 is built out of redudant blocks (which is ridiculous, of course), going from 6 down to 5 can only give you a theoretical maximum of 30% additional dies. And that for a part which would have an expected volume that's larger than it's more performant brother!

This would economically very inefficient: to meet this high volume, you'd have to sell working dies for less than they're capable of. Much smarter to just sell the higher performant ones with, say, slightly lower clock speeds to make it harder for the competition to catch up.

As Geo pointed out: All that doesn't exclude some lower end product down the road that will have 4 out of 8 active clusters for some smaller volume products, but to think that there ever were plans to make this the main initial product is very unlikely.

 

[hoom]

now rumoured 775MHz clock

That was my point with the 'r700 info' bit.

You can chuck any thing you want into the Diamond website & it gives an error saying that the page has been removed.
Make up a fake screenshot with specs you want people to believe & say it came from there:
Presto rumor

Unless anyone did actually see that page

 

[trinibwoy]

RV670 as the naming implies traditionally would point to an "value" sku. However all that changed when we heard that it was up and running ahead of schedule and it was effectively an improved (TBD) R600 and DX10.1. Now, 2900XT is very competitive with the current GTS which is a 96SP sku. Common sense would suggest that if they needed to combat an improved R600, they would need more SP than that .. Makes perfect sense.

Even if Nvidia thought RV670 was a "value" sku (which is highly doubtful) why would they target this high volume segment with a chip with half its units turned off? That would be a very expensive and low margin part. This is not a garbage SKU like the GTO's used to get rid of old inventory. This is a brand new chip. That's the point you don't seem to be getting.

Another thing to is the relatively conservatively clock on the GT. It will be interesting to see what the G92 GTS looks like.

 

[Lux_]

Unless anyone did actually see that page

Couple of pages back.

 

[ShaidarHaran]

Couple of pages back.

DB already debunked that. At least, I hope his dismissal was debunking and not just marketing FUD That or R680 and even R700 had better damn well be right around the corner (speaking purely from an enthusiast standpoint). Worst case IMHO is 2x Revival (at reported lower 668MHz clocks) on a PCB as 3890 or some such and no major improvements in R680 with R700 not launching until the 2nd Half next year.

 

[Arty]

Even if Nvidia thought RV670 was a "value" sku (which is highly doubtful) why would they target this high volume segment with a chip with half its units turned off? That would be a very expensive and low margin part. This is not a garbage SKU like the GTO's used to get rid of old inventory. This is a brand new chip. That's the point you don't seem to be getting.

Another thing to is the relatively conservatively clock on the GT. It will be interesting to see what the G92 GTS looks like.

Why is it highly doubtly? The fact that Nvidia's 8800 lineup is completely messed up is irrefutable proof right there. Besides, you dont need to repeat "half units disabled" when I cleared that their original plan was 96/128 (GT/GTS), which seems to be happening now (new GTS) according to the latest rumor.

Nvidia can sell a lot of 484mm2 dies not including NVIO (320MB GTS) in the $200-300 range, why cant they do a 330mm2 die (8800GT) in the $100-200 range?

The flexible configurabiity of G80 (G92) allows them to target a broader range of the market, in this case from the 8800GT upto the impending GX2.

 

[trinibwoy]

Why is it highly doubtly?

Simple. AMD had a target to beat in the 8800GTS-320. Why would Nvidia (or anyone else) assume that RV670 would be considerably slower than the GTS, especially since it has been known for some time that it's on 55nm? According to your logic, the new entrants to the $250 price segment (RV670 and G92) were both meant to be slower than the incumbent 8800GTS-320.

 

[Arty]

Simple. AMD had a target to beat in the 8800GTS-320. Why would Nvidia (or anyone else) assume that RV670 would be considerably slower than the GTS, especially since it has been known for some time that it's on 55nm? According to your logic, the new entrants to the $250 price segment (RV670 and G92) were both meant to be slower than the incumbent 8800GTS-320.

Simple. RV570 was 3/4 R580, RV670 on the same lines would be 3/4 R600 or a 240SP sku. That would be a nice step up from the 8600GTS/2600XT and with some high clocks could have matched the 320MB GTS while losing to a 96SP G92.

When was the last time AMD shrunk their flagship and called it a RV sku? This is the first and one that caught Nvidia offguard. However they managed to combat the situation albeit messing up their lineup.

 

[hoom]

Couple of pages back.

Not at all suspicious that the image is the same width & only a couple of pixels different in height than these linked by Fudo?
Notable differences being use of HD2600Pro boxes in the B3D post + some slight typo differences.