Roy Taylor: "Only Brits care about ATI"

Which is why sometimes it really sucks to be a designer.

Depends on how you look at it. You could also say that it sucks that some people have unreasonable expectations.
Look at them sitting at their XP machines, criticising nVidia for not having DX10.1 hardware yet, oh the irony.
 
Fleshing out the alternative history with the missing GPU, "GT200":
  • November GT200 launches in $500 + $600 variants. The G80 hangs in there at ~$400
  • January G94 launches at $250, the same price as G84 launched at (doh)
  • February/March as G80 runs out G92 arrives to take over in the $300/400 bracket
NVidia suffered two problems: the severe delay to GT200 (7 months+ it seems) and the early arrival of RV670.

It seems to me NVidia did exactly the right thing, bringing forward G92. It was a bit ropey at the beginning but it had the desired effect of making ATI's GPUs irrelevant for the majority of people who actually pause to consider what it is they're buying.

If NVidia had brought forward G94 instead they risked losing that marketing dominance for the most important quarter (not knowing RV670's performance). Indeed NVidia gave RV670 the benefit of the doubt, pitching against it a considerably better GPU than it eventually turned out needing to do. If that isn't a mark of NVidia competing with ATI ("caring"), then I don't know what is.

Jawed
 
Is the 7 months delay speculation, rumour, fact, ...?
There was definitely a chip planned for launch before Christmas last year (multiple independent statements relating to a double-precision capable GPU). The exact lateness will have to remain as speculation, much like R600's.

Jawed
 
I'm not aware of any roadmap in which GT200 was supposed to be available before G92 or G94. Furthermore, given a certain... something in GT200's design methodology, it would make no sense to release G9x derivatives after GT200's release (at roughly the same time, okay, that's different). And I honestly don't think G80 would have a chance in hell to hang in there at $400 with a GT200 SKU at $500 unless the latter massively underdelivers...

To me, it simply seems that G92/G98 were originally aimed at Summer 2008 and GT200 was aimed at the December/January/February. Just like as late as February (iirc) 2006, NV thought Summer 2006 would still be realistic for G80...
 
Of which year?
Oh, December 2007/January 2008/February 2008. I do have one non-public data point which indicates that to be the case, but it's anecdotal so don't consider this to be anything more than speculation... On the other hand, I do know with utmost certainty that for certain reasons, it wouldn't make any sense to release G9x chips more than 2-3 months after GT200 is out.
 
Oh, December 2007/January 2008/February 2008. I do have one non-public data point which indicates that to be the case, but it's anecdotal so don't consider this to be anything more than speculation... On the other hand, I do know with utmost certainty that for certain reasons, it wouldn't make any sense to release G9x chips more than 2-3 months after GT200 is out.
Which puzzles me, because you say that G92/98 were planned for summer 2008, ~6 months after GT200. I don't understand why you let this contradiction go.

Jawed
 
Aaargh, oops, that was either a brainfart or a typo, I'm not sure, but I definitely meant 2007 - sorry! :)
 
Aaargh, oops, that was either a brainfart or a typo, I'm not sure, but I definitely meant 2007 - sorry! :)
For a while I thought G92 was supposed to be a summer 2007 part. The 65nm problems at TSMC presumably hurt it as well as the ATI parts that released earlier.

The implication is that NVidia has spent something like 9 months spinning G92 to get it working right on 65nm (May to January, roughly).

Is G92 on its third spin?

Jawed
 
Is G92 on its third spin?
No, A12. First spin came back roughly in the timeframe you'd expect for a November launch IIRC, so I suspect it was just the tape-out that got delayed. Presumably the same is true for GT200 but who knows!
 
No, A12. First spin came back roughly in the timeframe you'd expect for a November launch IIRC, so I suspect it was just the tape-out that got delayed. Presumably the same is true for GT200 but who knows!
There's no doubt that GT200 would have been extremely unlikely to have been their first 65nm GPU, which indicates something else should have been launched before it.

I've got no idea what G98 was about but it seems strange to me that G92 ended-up as their first 65nm GPU.

As noted a while back, CUDA-specific documentation relevant to GT200 changed sometime in the middle of 2007, i.e. long after one would expect such changes to be made and still make it into a 2007-GPU. Left-hand v right-hand confusion or fall-out/opportunity from the delay? If 65nm problems at TSMC delayed GT200 this implies a heavy scheduling overlap with the other GPUs NVidia was also implementing on 65nm.

So, all in all, very risky. Seeing as GT200 is rumoured to be ~ as big as G80 it appears NVidia had no choice about using 65nm for it. So the question then becomes, were they actually trying to learn from other GPUs on 65nm or were they cornered into just going for it hammer-and-tongs?

Are there other 65nm NVidia GPUs from before G92 (I can't find any)? Were there any NVidia GPUs that were smaller than G84? I have this feeling NVidia has developed a recent habit of not releasing the smallest possible GPU for the lowest value SKUs. Hence the missing G98.

Jawed
 
Were there any NVidia GPUs that were smaller than G84?

G86 comes to mind... :LOL:

Also, there is something called "G96" (9500 GT), although it was just recently launched. Basically it's a die shrink from "G84" (8600 GT/GTS), only sold at lower "G86"-like (8500 GT) prices.
 
G86 comes to mind... :LOL:

Also, there is something called "G96" (9500 GT), although it was just recently launched. Basically it's a die shrink from "G84" (8600 GT/GTS), only sold at lower "G86"-like (8500 GT) prices.

It hasn't launched yet (June?). There's been some previews and it seems to be to the 8600 what the 3650 is to the 2600.
 
G86 comes to mind... :LOL:
:oops: What I was trying to ask is, what happened to the chip with only a 64-bit memory bus, for 8400/8300 or whatever they were going to be called. There seems to have been a chip that never appeared in the G8x range.

Also, there is something called "G96" (9500 GT), although it was just recently launched. Basically it's a die shrink from "G84" (8600 GT/GTS), only sold at lower "G86"-like (8500 GT) prices.
But no G98.

So for whatever reason the ultra-low-end discrete isn't worth making chips for as far as NVidia is concerned.

And what I was trying to get at was, what 65nm chips (GPUs or anything else) was NVidia working on before G92 launched?

It seems very strange that a high-end chip was their first 65nm chip, so I'm wondering if I've missed something.

Jawed
 
So for whatever reason the ultra-low-end discrete isn't worth making chips for as far as NVidia is concerned.
Quite on the contrary, they just screwed up.

And what I was trying to get at was, what 65nm chips (GPUs or anything else) was NVidia working on before G92 launched?
G78, just google it. But I'm not sure if it was ever released in the wild or not...
 
Quite on the contrary, they just screwed up.

G78, just google it. But I'm not sure if it was ever released in the wild or not...
It seems they've screwed up three in a row Gx8 GPUs :oops: Clearly, no one cares :LOL:

So, still the question hangs, NVidia was really going high-end on 65nm before they'd finished anything else? Either with G92 in summer 2007 (which I don't believe) simultaneously with G98, which got lost in the 65nm confusion. Or GT200 at the end of 2007 (if G92 and G94 were really 2008Q1 parts).

It all sounds bonkers.

The "missing G98" appears to slot in nicely into NVidia's Gx8 tradition. I can't help wondering if all the Gx8s were, in fact, smokescreens.

Or is the conclusion that 65nm was such a fuck-up that NVidia abandoned G78, "G88" and G98 (all of which were 65nm?) and just got stuck in with what really needed doing, G92 and GT200.

---

It's got me thinking that perhaps G80 was meant to be a May/June 2006 launch and G92 could have been scheduled to launch in summer 2007 replacing it. Perhaps with G84/86 launching for Christmas 2006 and G98 being a March 2007 launch on 65nm (i.e. about the same time as ATI was scheduling its first 65nm parts). G98's timing would then have prepared the 65nm ground for G92. This kind of schedule has some plausibility to it.

---

It's the simultaneity of G92 and G98 in Arun's timeline that makes me dubious about G92 being planned that early. Unless, as I say, all of G8x was running 4/5 months behind schedule - without that delay, G98 would have had an implementation window ahead of G92.

Jawed
 
It seems they've screwed up three in a row Gx8 GPUs :oops: Clearly, no one cares :LOL:
Haha. Well I doubt G78 was ever aimed at anything but OEMs anyway. Who knows, maybe it *is* available somewhere out there... I don't think anyone knows or cares enough to check, though. And I don't think G88 ever existed? I mean, that'd be what, 4 TMUs and 8 SPs? AFAICT they only added the half-octo-TMU flexibility in G9x... FWIW, I've got good reason to believe there will be a 4 TMUs/8 SP GT2xx design for IGPs (and no that won't be the only IGP design).

So, still the question hangs, NVidia was really going high-end on 65nm before they'd finished anything else? Either with G92 in summer 2007 (which I don't believe) simultaneously with G98, which got lost in the 65nm confusion. Or GT200 at the end of 2007 (if G92 and G94 were really 2008Q1 parts).
The plan was G78 first (and certainly *samples* were there first; whether they actually shipped it is another question and that depends on wafer cost, yields, etc. too) - and then G98 and G92 at the same time, yes.

The "missing G98" appears to slot in nicely into NVidia's Gx8 tradition. I can't help wondering if all the Gx8s were, in fact, smokescreens.
They weren't, they just screwed up bigtime with G98. I can't really point out the details, and I don't know all of them anyway...

It's the simultaneity of G92 and G98 in Arun's timeline that makes me dubious about G92 being planned that early. Unless, as I say, all of G8x was running 4/5 months behind schedule - without that delay, G98 would have had an implementation window ahead of G92.
Chips always run out of schedule, just some more than others. You'd fall off your chair if you realized the magnitude of some of the delays for certain chips at both NVIDIA and ATI in the last few years... And it's not just about GPUs either, quite on the contrary.
 
And I don't think G88 ever existed? I mean, that'd be what, 4 TMUs and 8 SPs?
That's why I put G88 in quotes - everyone was expecting something for the bottom of the barrel, with only a 64-bit MC, but it never appeared.

The plan was G78 first (and certainly *samples* were there first; whether they actually shipped it is another question and that depends on wafer cost, yields, etc. too) - and then G98 and G92 at the same time, yes.
Well that clearly contradicts NVidia's heretofore "don't risk high-end on a new process" maxim. Coinciding nicely with 65nm :p

Chips always run out of schedule, just some more than others. You'd fall off your chair if you realized the magnitude of some of the delays for certain chips at both NVIDIA and ATI in the last few years... And it's not just about GPUs either, quite on the contrary.
One way or another, they both seem to have made an awful lot of rubbish in the chipset arena. SB600 still appearing on new mobos? No thanks.

Jawed
 
Back
Top