NVIDIA GF100 & Friends speculation

[Tchock]

BTW regarding memory speeds wasn't there something mentioned about PHY interface portions being larger on Cypress to compensate?

Again could also be the engine vs memory clock tradeoff for optimal perf/W instead of paper tiger spec sheets...

 

[3dcgi]

Unless of course that gap in the balls means nothing

It means nothing. I've looked at a few chips and there's no correlation between the size of the inner square and the die. Also, some ATI dies are rotated on the package yet the pins are not so it's obvious from this example that there's not a direct correlation of die to pins.

 

[rjc]

If you go by the right hand side then I'd count 10 layers. And, no, 5800 series are not "quite a few more".

Yeah sorry looked again at picture as soon as i posted it... My memory is telling me the orginal GTX280 board had 14 and with the GTX285 they reduced it to 10 for the reference deisgn....is that correct?

 

[mczak]

The next chip is apparently in process and is slightly larger than the 65nm g92, not the 55nm g92b.

Really? G92 was as large as Cypress. That can't be good.

There was a widepread rumor of a 256shader/256bit part, but it appears that for some reason those specs have been cut down.

So the larger-than-g92 chip isn't even 256shader/256bit mem interface? That doesn't sound right. AMD got a half-cypress part at roughly 55% die size of cypress. But larger than g92 would be ~70% of the rumoured 500mm^2 GF100, and it doesn't fit half the shaders (though admittedly that would be more than half the memory bandwidth and thus likely rops too)?

 

[rjc]

Really? G92 was as large as Cypress. That can't be good.

Quote from cncfc(post #172) after someone else comes out about the "GF104's" diesize:

我问了下，说是GF104比 65nm的G92略大，不是双256，这年头打听消息越来越难
(I to ask downstairs, yes GF104 near to 65nm of G92 in plan size, is not pair of 256, this year to enquire about information is more and more difficult)

Slightly contradicting, others in the thread claim the part is 'pair of 256' also performance round 5830-5850 and priced below GTX285 level.

So the larger-than-g92 chip isn't even 256shader/256bit mem interface? That doesn't sound right. AMD got a half-cypress part at roughly 55% die size of cypress. But larger than g92 would be ~70% of the rumoured 500mm^2 GF100, and it doesn't fit half the shaders (though admittedly that would be more than half the memory bandwidth and thus likely rops too)?

The die size seems to be fairly concrete, other information which pointing to around that expected performance and pricepoint is maybe not quite as strong. Are you absolutely sure the GF100 is only around that die size?

 

[Tchock]

不是双256

could mean 256bit + 384 ALUs though. Not really any indication of anything.

 

[FenderBender]

Sorry if these were posted before... they were new to me anyway.

 

[rjc]

不是双256

could mean 256bit + 384 ALUs though. Not really any indication of anything.

Negative logic though - he is eliminating things isn't he? ie the 256shader/256bit specification

A few pages before:

这倒是给了NV时间，Fermi说实在的从GF100就看出来die size控制得不好，腾出时间仔细优化一下。
(This to know NV time, Fermi to speak honestly of unhurried GF100 only to see die size control obtained is not good, to rise time spent optimizing is needed)

ie if have translated correctly the before he is saying nvidia should spend time optimizing their die size.

There is another post yesterday in a dffierent thread:

GTX480没有512SP，产量确实很少，都给Tesla了
NV现在重点已经是GF104了，GF104在实验室里性能已经接近GF100
(GTX480 is not to be 512SP, output indeed very few, all to supply Tesla.
NV's focal point now already is GF104. GF104 in internal laboratory is close to GF100)

ie GTX480 has been cut down,as others saying probably 480shaders. Cannot supply market with sufficent product to meet demand therefore working on chip below which they think will get close in performance to GF100 once it tapes.

 

[Ailuros]

Above G92@65nm die area chimes in well with what I heard about it some time ago. As far as the unit count/memory interface goes I'd place my bets on two variants from the same chip and not just one and yes that might be very well a point all this confusion comes from.

 

[rpg.314]

IF 470 is somewhere between a 5850 and a 5870, then is this GF104 @330mm2 thing going to compete with Juniper @181mm2 (or 5830 for that matter)? Is the entire architecture fucked up?

 

[Blazkowicz]

GF104 and GTX 470 may feel a lot like geforce 6600GT vs 6800 - the 6800 was much more beefy and had much slower clocks, making the cards almost equal with a very slight edge for the 6800.

I wonder if the GF104 is a version without DP, or maybe DP running quarter rate instead of half rate.

probably GF104 is a good contender against 5850, it can target both Cypress and the "missing GPU" between Juniper and Cypress.
(speaking out of my buttocks, hope the smell is ok)

 

[Ailuros]

I'd suggest we wait for independent measurings from serious websites after the launch. I don't trust any supposed results during ultra-silly-season however good or bad they may sound.

 

[Ailuros]

GF104 and GTX 470 may feel a lot like geforce 6600GT vs 6800 - the 6800 was much more beefy and had much slower clocks, making the cards almost equal with a very slight edge for the 6800.

I wonder if the GF104 is a version without DP, or maybe DP running quarter rate instead of half rate.

probably GF104 is a good contender against 5850, it can target both Cypress and the "missing GPU" between Juniper and Cypress.
(speaking out of my buttocks, hope the smell is ok)

GF104 sounds more like half a GF100 if all indications are correct. It shouldn't have anything to do with a 470 which is obviously a castrated GF100 chip.

The plain 6800 and all other 6800s for that matter came from the same chip.

 

[Blazkowicz]

that precisely is the way I look at it - 6800 the castrated chip, 6600GT the half-spec chip that goes high on clock.

 

[Ailuros]

that precisely is the way I look at it - 6800 the castrated chip, 6600GT the half-spec chip that goes high on clock.

With one eye closed (NV40 had 3 and not 2 variants at launch) it could work. Officially announced so far we have a GTX480 and a GTX470. Considering how low the specificiations for the latter already seem to be I doubt they will kick off another (<470) variant from the same chip at launch.

GF104 isn't slated for March anyway afaik.

 

[ZerazaX]

不是双256

could mean 256bit + 384 ALUs though. Not really any indication of anything.

I can read enough Chinese that that statement says:

Is not pair of 256

So not 256/256 is what he wrote

 

[leoneazzurro]

IIRC (and I should do as I had one) the 6800 came in several "flavours" at launch (6800 Ultra, GT and vanilla) and it's true that the plain 6800 was not much faster than the 6600GT (except for corner cases, i.e. bandwidth limited, but also 6600 GT came some time after 6800's launch) but at that time you could also try to re-enable "sleeping" pipelines and the overclock usually was very good on that parts. And there were some 6800 GT - 128 Mbyte parts (standard parts came with 256 meg) which costed not so much more than the 6600 GT but packed a lot more performance (and not considering that almost all reached 400+MHz on the core without modifications or strange cooling systems)...
Of course, it is possible that the GF104 can compete successfully with the 5850 if the clocks are in the 1,5+ GHz range or if it is a 384 SP part (but I wonder then what the die size could be).

 

[dizietsma]

If the performance figures are true then it's not perked my interest at all. Reminds me of the R520 -> R580, all those extra pixel pipelines were a bit of a waste, I wonder if this is going to be same, with tess. hardware not really adding much for most things.

 

[Rys]

Which it can. Packages only dictate max die sizes. Its likely that gap only means that they went with outside signal arrays and a dense center P/G array for the balls.

It means nothing. I've looked at a few chips and there's no correlation between the size of the inner square and the die. Also, some ATI dies are rotated on the package yet the pins are not so it's obvious from this example that there's not a direct correlation of die to pins.

I'd always gone with the outer ring being for power and ground plane(s), with the inner block for I/O, but with the gap still marking the chip edge. Throw out ~500mm/sq again then I guess, you both make good points.

 

[NathansFortune]

So what are we saying, smaller than 500mm2?

That seems awfully small after everything we seem to have learned from it...

If it comes in at around 22x22mm that would be quite a surprise no? I was expecting something closer to 24x24mm.