ELSA hints GT206 and GT212

[DegustatoR]

Funny.
Few years ago someone was claiming that altho ATi has low and mid cards on smaller chips and process, still NV can compete because of using proven less expensive process.

Do you see any difference between "now" and "eventually"?

 

[Arun]

Nvidia has no competitive parts at all right now. Zero. Their die size make it impossible to compete.

Charlie, you're obviously right that NVIDIA is still at a disadvantage here. Mind you, their disadvantage is much smaller than AMD's was in the R5xx timeframe or what NVIDIA had in the NV3x timeframe, so the financial consequences should not be exagerrated (of course, you could argue that NVIDIA would have been hit much harder were it not for NV33, aka the chip nobody ever heard about on the quarter-node process that barely exists but still generated the majority of NV's revenues in 2004 - still let's not get ahead of ourselves).

However, you are ignoring three things. First of all, GT216/GT2178 are more competitive in the low-end than NV's previous chips would have been for the Winter cycle and AMD hasn't refreshed that part of their portfolio yet, so NV's situation actually improved there in the last 6 months (even if it's still far from ideal to say the least)

Secondly, Fermi derivatives will come faster than you think (no, they haven't taped-out yet) and they'll be slightly more area-efficient than GF100. GT215 will probably turn out to be a fairly short-lived chip. And yes, free to come back to this post in a couple of months if I'm wrong; I'm all for accountability.

Thirdly, stopping GT200 production was a conscious move because you are right: they couldn't sell them without losing money anymore. They didn't want to repeat the fiasco that happened when RV670/G92 launched and they had too much G80 inventory. That made them lose truckloads of money. It's bad that they are in such a situation but amongst the awful strategic choices available to them, it's still (amongst) the best.

 

[mczak]

(GT215s (~125mm²) die-size competitor is clearly RV740

Oh that small? I thought it would be bigger (merely because GT220 is already ~100mm²). Well I guess in this case it's quite a ok chip for the die size, not really surprising it can't compete with rv740 then, maybe it'll be priced accordingly eventually.

 

[DegustatoR]

Oh that small? I thought it would be bigger (merely because GT220 is already ~100mm²). Well I guess in this case it's quite a ok chip for the die size, not really surprising it can't compete with rv740 then, maybe it'll be priced accordingly eventually.

G92 is able to compete with 4850, why do you think that GT215 won't be able to compete with RV740?
Another question here is what will Redwood be as that chip looks to be the one that will compete with GT215 for some time. If that's half-Juniper then that's 400 SPs/20 TUs against GT215 which should be able to reach 640 SPs/32 TUs level of RV740. Since Redwood will retain 128-bit bus (but loose GDDR5 support?) it will probably be less of a jump in die size than that between 256-bit Cypress (334?) and 128-bit Juniper (~180?). That puts it around 100-120mm^2 while having substantially less impressive specs than RV740 (which was 137mm^2) which should be the top level of GT215 performance.
I don't really see what's disasterous are there for GT21xs in light of these estimations. NV may have to cut their prices a little (GT216 should be at ~$50, GT215 at ~$80) to adjust to market conditions but that they probably will be able to do as 40G matures and TSMC's prices go down. What should be undeniable here is that all the GT21x parts will eventually cost less to produce than previous G9x parts and AMD's current RV7x0 low-cost parts. And while their timing / performance / features compared to competition may be bad I don't really see how their introduction instead of G9x GPUs and against EG 40G low-end GPUs is bad for NV in the long run.

 

[AnarchX]

G92 is able to compete with 4850, why do you think that GT215 won't be able to compete with RV740?

G92: 16 ROPs, 64 TMUs, 128SPs@ up to 1.8GHz
GT215: 8 ROPs (a bit improved), 32 TMUs, 96SPs(a bit improved)@ up to 1.6GHz?

And if you believe charlie NVs GDDR5 efficiency is a disaster...

 

[mboeller]

and 128-bit Juniper (~180?).

according to anandtech Juniper is around 166mm². So half of that would mean Redwood is <100mm²

 

[DegustatoR]

GT215: 8 ROPs (a bit improved)

It supports GDDR5. Why would they support GDDR5 an not have enough ROPs to utilise it?

96SPs(a bit improved)@ up to 1.6GHz?

The whole GT21x clocks thing is a bit mysterious to me right now because most of cards overclocks like hell (+30% and more than 1 GHz for ROPs in some cases). I don't understand why would NV set such low reference frequencies.

And if you believe charlie NVs GDDR5 efficiency is a disaster...

Thanks, but I tend not to believe anything that comes from a PR department.

according to anandtech Juniper is around 166mm². So half of that would mean Redwood is <100mm²

And according to Hexus authors who actually have it measured it's 12x15=180 mm^2.

 

[mczak]

G92 is able to compete with 4850, why do you think that GT215 won't be able to compete with RV740?

Because the hd4770 is nearly as fast as hd4850, while clearly these early benchmarks indicate gt215 isn't close to gts 250 even in its presumably fastest configuration and overclocked. Though that's only 3dmark for now, actual games remain to be seen, but I'd be surprised if things would change a lot. Clocks are suspiciously low though so maybe there's some room for improvements. Memory isn't exactly very fast neither, but since it appears to have only 8 rops faster memory might actually not help (the difference between the gddr5 version and the gddr3 version (with half the bandwidth) is already quite small).

 

[Jawed]

And according to Hexus authors who actually have it measured it's 12x15=180 mm^2.

Measuring the packaged die is not the same as measuring the chip. 0.5mm of packaging appears to be the norm, 11.5x14.5mm = 167mm².

Jawed

 

[mboeller]

And according to Hexus authors who actually have it measured it's 12x15=180 mm^2.

and I think they made a mistake.

If you compare the 5 pence (18mm diameter) with the RV790 die, the RV790 die would have a area of ~ 305 mm². Using the same measurement the RV840 die has said ~180 mm².

Therefore IMHO they didn't substracte the outer edge from the measurement when they said that the RV840 has 180 mm². Removing this outer edge the die of the RV840 is 166 mm²

So 166 mm² is imho the correct die size of the RV840.

 

[trinibwoy]

G92: 16 ROPs, 64 TMUs, 128SPs@ up to 1.8GHz
GT215: 8 ROPs (a bit improved), 32 TMUs, 96SPs(a bit improved)@ up to 1.6GHz?

GT216 already has 8 ROPs, what makes you think GT215 won't improve on that? Even so, I doubt it will catch G92 even with 16.

 

[DegustatoR]

Jawed, mboeller, does the same rule apply to GT218/216/215 die measurements? If yes then you have to adjust them by 0.5x0.5 too?

 

[Jawed]

Jawed, mboeller, does the same rule apply to GT218/216/215 die measurements? If yes then you have to adjust them by 0.5x0.5 too?

Should do - depends on whether the dimensions are as stated by NVidia or as measured by a reviewer including the packaging.

Jawed

 

[Dave Baumann]

So 166 mm² is imho the correct die size of the RV840.

Turns out that its somewhere inbetween - the 166mm² quote was actually a pre-silicon number; the final "official" size is 170mm².

 

[elsence]

It seem to be only 8 ROPs, else it would not be slower than 9600 GT.


With 55nm wafers @ ~80% costs of 40nm ones and probably higher yields on 55n, I see no real advantage for:
GT216 100mm² vs RV730 140mm²
GT218 57mm² vs RV710 71mm²
(GT215s (~125mm²) die-size competitor is clearly RV740(


According to Anandtech Juniper is 166mm², so Redwood @~100mm² could be a 400SPs part and Cedar migth have the chance to be a bit faster than RV710 while be in GT218s range.

When i first saw the 128bit memory controller i said 8ROPs based on previous desktop DX10 NV architectures.
But if you check the Vantage/3DMark06 scores, a 8ROPs design can't have this performance.
I mean, if the design is like you say 8ROPs/32TUs we have:

GT240 = 8ROPs / 32 TUs / 96SPs / 550MHz core / 1340MHz shaders / 900MHz GDDR5 / 57,6GB/s bandwidth

GT250 = 16ROPs / 64 TUs / 128SPs 700MHz core / 1725MHz shaders / 1000MHz GDDR3 / 64GB/s bandwidth

How is it possible a part like GT250 with 2,5X pixel fillrate with 2,5X texel fillrate and with 1,7X shading speed to be only +6% faster? (5983 vs 5613)

In the mobile space Nvidia had some designs with 8ROPs per 64bit channel, i suspect this will happen here also.

In Vantage the 550MHz GT240 (GDDR5) is around 43% faster than 600MHz 9600GT. (5613 vs 3899)

In 3DMark06 (1280X1024) the 550MHz GT240 (GDDR5) is around 2% faster than 600MHz 9600GT. (9478 vs 9224)

In 3DMark06 (1920X1200) the 550MHz GT240 (GDDR5) is around 8% slower than 600MHz 9600GT. (7629 vs 8313)

I suppose is the memory controller paired with GDDR5 (efficiency, latency, etc...)

But I don't like the Vantage scores (GT240 vs GT250) something is fishy...

If the GT216 is 486 million transistors 100mm2 die size, the GT215 should be at least 725 million transistors which means at least 137mm2 die size at 40nm (i forecast 780 million transistors with 144mm2 die size)

Regarding performance:
Redwood>RV730>GT220
Cedar>RV710>G 210

Even in the case that Redwood is a DX11 RV730 with 64bit memory controller and GDDR5 (lowest possible scenario...) it will be faster than GT220 and the die size would be 93-100mm2.

I 'll post later my predictions.

*EDIT*
Right now the 5750 512MB cost 110$.
GT 220 with 512MB GDDR3 cost 80$.
What price the GT240 DDR3 and GT240 GDDR5 will have?
At least the GT240 GDDR5 will fight against 5750 and will lose handily.

 

[Mintmaster]

G92 is able to compete with 4850, why do you think that GT215 won't be able to compete with RV740?

Another question here is what will Redwood be as that chip looks to be the one that will compete with GT215 for some time. If that's half-Juniper then that's 400 SPs/20 TUs against GT215 which should be able to reach 640 SPs/32 TUs level of RV740.

G92 gets close to the 4850 for the same reason that it competes with the GTX 260. The latter two are cut down parts with lower clocks and/or fewer SPs. RV740 is a bit faster than a GTS250, which is substantially faster than GT215, judging by the info that has leaked out so far.

Redwood probably will compete with GT215 a bit (maybe the lower clocked ones), but only because AMD will choose higher margins rather than try to make NVidia bleed. Looking at GT216 and GT218 pricing, NVidia would rather have low sales than zero margins. AMD is in no position to turn down higher ASPs, so my guess is that Redwood and Cedar will have higher prices than we'd like.

What should be undeniable here is that all the GT21x parts will eventually cost less to produce than previous G9x parts and AMD's current RV7x0 low-cost parts. And while their timing / performance / features compared to competition may be bad I don't really see how their introduction instead of G9x GPUs and against EG 40G low-end GPUs is bad for NV in the long run.

'Eventually' is a big problem, because right now ATI has successful parts that are not only mature, but are also going to be discounted when going EOL to make way for new Evergreens.

Remember that checkmark features and the Halo effect matter a lot for the value market, too.

 

[Mintmaster]

But I don't like the Vantage scores (GT240 vs GT250) something is fishy...

Something about Vantage has always given GT2xx a marked advantage over G9x. It may be that the shaders need more registers, thus reducing the latency hiding ability and efficiency of G9x.

 

[elsence]

Something about Vantage has always given GT2xx a marked advantage over G9x. It may be that the shaders need more registers, thus reducing the latency hiding ability and efficiency of G9x.

Probably, but this difference is huge.

If we compare:

a GTX260 216 (1242MHz 216SP)
or a GTX275 (1404MHz 240SP)
with a GTS250 (1836MHz 128SP)

in 3DMark Vantage (Performance) scores we will see that the GT2XX has only +15% advantage per MHz*SP.

http://www.xbitlabs.com/articles/video/display/mainstream-roundup-2_13.html#sect2

In the it168.com 3DMark Vantage (Performance) tests we will see that:

a GTS250 (1725MHz 128SP) with 64GB/s bandwidth is
only 26% faster than
a GT240 DDR3 (1340MHz 96SP) with 28.8GB/s bandwidth.

So in this case the advantage is +35% per MHz*SP instead of the usual +15%.

If we compare with a GT240 GDDR5 (1340MHz 96SP) with 57.6GB/s bandwidth the advantage would be +60%.

Something is fishy...

 

[Groo The Wanderer]

However, you are ignoring three things. First of all, GT216/GT2178 are more competitive in the low-end than NV's previous chips would have been for the Winter cycle and AMD hasn't refreshed that part of their portfolio yet, so NV's situation actually improved there in the last 6 months (even if it's still far from ideal to say the least)

More competitive usually means less subsidies. You are right, it buys them a bit of breathing room for three months. Then again, if you look at the margins on these parts on a good day, it isn't enough to move the financial needle much. The segment is high volume, crap margins.

So yes, definitely improved. Above or below the profit line is still an open question.

Secondly, Fermi derivatives will come faster than you think (no, they haven't taped-out yet) and they'll be slightly more area-efficient than GF100. GT215 will probably turn out to be a fairly short-lived chip. And yes, free to come back to this post in a couple of months if I'm wrong; I'm all for accountability.

I know when they are planned, and I know why they haven't taped out yet. In fact, I think I was the first (only?) one to say it.
http://www.semiaccurate.com/2009/08/05/four-more-gt300-variants-tip/
The thing is, 'slightly' more area efficient doesn't cut it. You need to be vastly more efficient to have a chance. Then again, depending on what they cut, they could be hamstringing their own message. Watch the PR monkey dance!

Thirdly, stopping GT200 production was a conscious move because you are right: they couldn't sell them without losing money anymore. They didn't want to repeat the fiasco that happened when RV670/G92 launched and they had too much G80 inventory. That made them lose truckloads of money. It's bad that they are in such a situation but amongst the awful strategic choices available to them, it's still (amongst) the best.

I know it was a conscious move, and I wrote that, before anyone else. I also wrote up the reasons why in great detail. The thing is, with the G80, they had an imminent replacement. GF100-lite(s) are not due for minimum 3-4 months, and that is _IF_ they figure out what is wrong with Fermi the Elder.

The thing that Nvidia hasn't figured out yet is that they are making themselves look like complete fools. Right now Jen-Hsun "puppy wrangler" Huang can't look dumber. They didn't send their partners the usual EOL letters because they didn't want Wall Street to find out about their chances for Q4. I talk to a lot of them, and the spreadsheet revision has been heavy to say the least. Guess how many partners bought PCBs and components for the christmas season in anticipation of normal GPU shipments? Anyone got a use for a lot of 2xx PCBs with no chips?

Then comes the chipset fiasco. Can you look dumber than by splitting hairs over not making any more chipsets vs shutting down the chipset division? Come on, that distinction won't fly on a third grader, but sadly, many tech sites puked it back in a story. SIGH. Then there was the claim that they are stopping development because of evil Intel. Intel is _SO_ evil that they stopped AMD chipset development too!

Nvidia doesn't have a single competitive part. ATI is in control top to bottom for at least the next 3 months, likely 6. They have a cheaper part, so they can run NV margins to whatever point they feel like, and NV can't so squat. NV screwed their partners over for PR reasons, and in general, can't do anything dumber. Just wait, I will probably eat those words in a few days.

So, better position? Yeah. Which refugee camp in Dharfur would you like to call home?

-Charlie

 

[Mintmaster]

Probably, but this difference is huge.

If we compare:

a GTX260 216 (1242MHz 216SP)
or a GTX275 (1404MHz 240SP)
with a GTS250 (1836MHz 128SP)

in 3DMark Vantage (Performance) scores we will see that the GT2XX has only +15% advantage per MHz*SP.

I doubt that's the right metric to use. 3DMark Vantage isn't solely SP limited - not by a long shot.

What I was trying to point out is that the GTX260 is 1.05x the speed of the GTS250 in 3DMark06 (using your link), but 1.37x the speed in Vantage. This jives perfectly with with the GT240 being 0.69-0.74x the speed of the GTS250 in 3DMark06 but 0.94x in Vantage.

There's nothing really fishy about it (or conversely it's been fishy ever since GT200 was released).