Nvidia GT300 core: Speculation

[rjc]

My guess is that AMD is going to get a 128-bit version of RV770 out there - with faster GDDR5 it should be fine. The aim being margins more than market positioning.

That's the RV740, according to the link on my previous post, Over 1 million have been produced....going to try to wipe out everything from high end 9500s up to the 4850.

It'd be nice if there was a monster RV790 coming out, at roughly the same die size on 40nm. It just seems way too quiet on that front though.

The RV790 is just a RV770 on a new 55GT process, coming in a few weeks.

Nvidia are not stupid, why push the GT212 at all which as it will only sell a small quantity when the 740 is just about to take a big swipe at your volume markets?

 

[INKster]

Nvidia are not stupid, why push the GT212 at all which as it will only sell a small quantity when the 740 is just about to take a big swipe at your volume markets?

That's where the GT214, GT215, GT216 and GT218 come in...
In fact, GT214 might just be a serious foe for RV740.

 

[turtle]

So does anyone else think the next battle of the titans will look somewhat like this?

GT212:

289mm2 (17x17)
384sp
96 TMU (67.2 bt/s)
16 ROP
256-bit
700/1750/6250 (2TF, 200Gbp/s)
12 TCP (32sp, 8 TMU per cluster)
>150W

GT: 288sp/72 TMU/ 12 ROP/ 192-bit/9 TCP/1.5TF/150Gbp/s
<150W

Rv870:

196-225mm2 (14x14-15x15)
400x5 = 2000sp
80 TMU (60 bt/s)
16 ROP
256-bit
750/6250 (3TF, 200 Gbp/s)
10 Arrays (200sp, 8 TMU per array)
~150W

5850: 625/4700 (2.5TF/150Gbp/s)
<150W


That would be nice to see, but maybe it's just wishful thinking on my part. Up until the 'news' broke about gt212, I thought those specs were more fitting for GT300 (on a larger die). I was thinking GT200 shrink + 384-bit for GT212.

As for 740: I still think:

121-144mm2 (11x11-12x12)
128-bit
640sp
32 TMU
775/782mhz (1TF) + 4/5gbps GDDR5- XT
625/xxxx (800Gflops) + GDDR3/5 - pro
I think the slower part would be 625mhz (800Gflops) and use GDDR3, although surely 4gbps GDDR5 would be sweet to see if the XT uses something faster. Comparing possible specs, a "pro" could take on my prediction of GT216 if equipped with GDDR5. There's got to be a reason Asus ordered a million of them (according to a post on chiphell).

That speculation would be:

GT216:

169mm2(13x13)
12 ROP
160sp
32 TMU
192-bit
630c/1575s/2400m (750 Gflops, 57.6Gbp/s)
768MB

GT215: Some butchered form of the above.


As for 790, I think it's a stop-gap of sorts. Theoretically a rv740 should start slower than a 4850 (barely) and overclocked could possibly get too close for comfort on 4870. I think 790, be it through more voltage/higher clock speed (as speculated) or rather the addition of more sp's (I still think [the?] two redundant arrays on rv770 will be turned on...voila 960sp/48 tmu part) it needs to be faster than rv770 because such parts are about to become essentially redundant. If you look at it as a 960/48 part, It's very possible they could clock such parts @ 1.5 and 1.25TF (782/651), putting a nice 50% performance disparity in-between it and rv740, and a 25% over rv770 (there-by giving reason for people to by the bigger 55nm die over the smaller 740)...all the while treading water waiting for 870, which would be essentially double it's performance. Of course, the same would go for rv770 parts clocked higher...938/782...but 900+ mhz doesn't seem likely. Like Grey Poupon, anything less would seem uncivilized.

Not to forget GT214, I think it's plausible it will end up like this:

110-121mm2 (11x11)
8 ROP
64sp
16 TMU
128-bit
520c/1300s/ (250 Gflops)
256/512MB


Well, with 790/740 supposedly coming Februaryish, I suppose it won't be long until we know for sure.

 

[rjc]

That's where the GT214, GT215, GT216 and GT218 come in...
In fact, GT214 might just be a serious for for RV740.

Yes but vr-zone had them listed as Q3 parts, and didnt even mention the GT215. Actually looks like nvidia's plan from 6 months ago, not today.

The GT215 was a surprise, it implies 5 chips for the desktop market. This seems too many now - the top of the market the high end has shrivelled and the bottom appears to be facing intense competition from igps and people abandoning desktops for notebooks and netbooks. Its sort of crunched up in the middle more, 5 chips is way too many for this, can get by with much less now than before.

Am thinking now the GT215 replaced the GT216 or pushed it down to entry level and the GT218 is gone. The GT212 is such an easy chip to delay as it has no real direct competition, and the current shrinking market is serviced by the just introduced GT200b. As long as the GT214 doesnt get anywhere near the 55nm GT200 everything will be fine. That and the GT200b appears to have taken 4 revisions to get working on 55nm suggests caution before trying to shrink it again.

 

[turtle]

I agree with most of that. I think that either one of those chips was pushed out, or they are codenames for parts rather than chips. IE 218 is a cut-down 216 and 215 of 214. If 218 does exist as a stand-alone chip, it's probably just ANOTHER iteration (40nm) of the 8600/9500...The bare minimum. Think g98 (8400gs)

As for the 55nm GT206, I think it will have a niche of it's own (between the 160sp 214 and 384sp 212) until they can finally phase the monster out...Which I imagine is as quickly as they can.

I think GT212 is 870's competition; I agree they don't need to hurry to get it out unless they THAT are worried about rv790 vs Gt206.

214 though, is incredibly important imho (as these parts always are for nvidia; they seem to last more than one generation like the 6600gt/7600gt/9600gt). They need a smaller chip to compete with rv770 ASAP.

 

[DegustatoR]

For the sake of argument, AMD could drop the CrossFireX Sideport and get a 256-bit bus on a ~200mm2 die (GDDR5 takes more pads, so slightly bigger than for a GDDR3 die).

60 mm^2 for what is essentially a PCIE 2.0 x16 interface? I think you're overestimating its complexity.
Plus RV770 on 40nm should be quite smaller than 200 mm^2 (i'm getting ~120 mm^2). Such chip will end up between RV770 and RV870 in performance (something like 1600 SPs / 80 TUs). It's possible, sure, but probably not in 2nd Q if RV870 is supposed to hit retail in 4th Q.

 

[Jawed]

60 mm^2 for what is essentially a PCIE 2.0 x16 interface? I think you're overestimating its complexity.

It seem it's perimeter that's the constraint, as I wrote here:

http://forum.beyond3d.com/showpost.php?p=1249850&postcount=110
http://forum.beyond3d.com/showpost.php?p=1249868&postcount=113

Plus RV770 on 40nm should be quite smaller than 200 mm^2 (i'm getting ~120 mm^2).

Agreed, quite small and only workable with a 128-bit interface.

Such chip will end up between RV770 and RV870 in performance (something like 1600 SPs / 80 TUs). It's possible, sure, but probably not in 2nd Q if RV870 is supposed to hit retail in 4th Q.

Trinibwoy was suggesting a "monster" RV790 and I was suggesting that without CrossFireX Sideport (which necessarily increases die area quite considerably, due to perimeter constraints) something around 200mm2 might work out, i.e. a little bigger than RV670 which has 256-bit GDDR3 - GDDR5 takes more die area too.

There's not been a real sniff of "RV790" so far, so I still think NVidia's going to have spring/summer/autumn to itself. AMD's made a song and dance about the "sweet spot" chip and its ability to introduce a complete top-to-bottom line-up very quickly, but nothing about what it means for refresh...

Jawed

 

[Vincent]

It is rumored to be the news that AMD might launch RV790 and RV770 in the spring of 2009. (http://bbs.chiphell.com/viewthread.php?tid=37831&extra=page=1)




Btw

My purchase consideration for new installation of my video card :

1. Cost less than 100 pounds

2. GT200 level performance

3. Used card only


Although it seems very unlikely this year

 

[rpg.314]

There's not been a real sniff of "RV790" so far, so I still think NVidia's going to have spring/summer/autumn to itself. AMD's made a song and dance about the "sweet spot" chip and its ability to introduce a complete top-to-bottom line-up very quickly, but nothing about what it means for refresh...

I find the GT212's Q2 launch rumours fake. Designing chips takes 2-3 years. Nvidia must have known that RV770 will win big despite GTX295. And why launch it? It's only DX10 and has ~6 months on the market before DX11 parts take over. It doesn't sound plausible to me.

 

[DegustatoR]

I find the GT212's Q2 launch rumours fake. Designing chips takes 2-3 years. Nvidia must have known that RV770 will win big despite GTX295. And why launch it? It's only DX10 and has ~6 months on the market before DX11 parts take over. It doesn't sound plausible to me.

You should view GT212 as a GT200 shrink. (It'll have more shading power because GT200 kinda sucked compared to RV770 so it's natural that they'll try to improve it.) NV done this with G71, G92 etc. GT300 will probably be to GT212 what GT200 was to G92b -- only this time it'll be more like G80 vs G71 since GT300 will most definitely have a refreshed architecture with DX11 support.
So 2Q08 for GT212 is pretty much expected for a long time. Wether they'll be able to hit this target is another matter...

 

[Jawed]

I find the GT212's Q2 launch rumours fake. Designing chips takes 2-3 years. Nvidia must have known that RV770 will win big despite GTX295. And why launch it? It's only DX10 and has ~6 months on the market before DX11 parts take over. It doesn't sound plausible to me.

GT2xx seems to be 6 months behind schedule generally...

Jawed

 

[rpg.314]

You should view GT212 as a GT200 shrink. (It'll have more shading power because GT200 kinda sucked compared to RV770 so it's natural that they'll try to improve it.) NV done this with G71, G92 etc. GT300 will probably be to GT212 what GT200 was to G92b -- only this time it'll be more like G80 vs G71 since GT300 will most definitely have a refreshed architecture with DX11 support.
So 2Q08 for GT212 is pretty much expected for a long time. Wether they'll be able to hit this target is another matter...

I don't get this. In the AT piece on RV770's story, they said that it takes 2-3 yrs to bring a chip to market. Now, with this chip, nV seems to be migrating to a new process, new RAM, and packing in more ALU's in ~1yr. If it is really so simple, why did rv770 take so long? Or is it because there are no architectural changes here and ALU's are being copied from GT200?

 

[Lukfi]

It takes longer to design a new architecture than to develop a new chip on the existing arch. and modifying it for a new manufacturing process. RV770 development started prior to R600 release, but it's very possible that at that time it was designed as a 96 ALU chip.

 

[pjbliverpool]

If GT300 is coming out this year then I think I'll try and hold out for it. It should have enough of a lead on the next gen consoles to "feel powerful" for a good couple of years while its not so far away that my 8800GTS will become useless in the meantime.

I'm keen for something that will play Crysis and GTA4 better (plus a handfull of other games that could do with more image quality/framerate) but there's nothing the GTS can;t at least get by in. A GTS version GT212 sounds tempting but thats gonna feel old pretty quickly once GT300 arrives with a new architecture and DX11.

 

[INKster]

I find the GT212's Q2 launch rumours fake. Designing chips takes 2-3 years. Nvidia must have known that RV770 will win big despite GTX295. And why launch it? It's only DX10 and has ~6 months on the market before DX11 parts take over. It doesn't sound plausible to me.

G71 was a thorough redesign of G70 for a brand new process (90nm) in a much smaller core, was on the market as a high-end for a mere 6 months, and precluded the process that was to be used on the "real" next generation, the G80, a very large chip by any standards.

What makes you think all of this can't happen again ?

 

[rpg.314]

G71 was a thorough redesign of G70 for a brand new process (90nm) in a much smaller core, was on the market as a high-end for a mere 6 months, and precluded the process that was to be used on the "real" next generation, the G80, a very large chip by any standards.

What makes you think all of this can't happen again ?

I didn't know that. Been Following GPU's since G80.

 

[3dcgi]

I don't get this. In the AT piece on RV770's story, they said that it takes 2-3 yrs to bring a chip to market. Now, with this chip, nV seems to be migrating to a new process, new RAM, and packing in more ALU's in ~1yr. If it is really so simple, why did rv770 take so long? Or is it because there are no architectural changes here and ALU's are being copied from GT200?

The 2-3 years mentioned involves some overlap between projects. Commonly high level architects move to the next design before the current design is ready to go. Plus there is a period of time between tape out and retail where most of the design team has started working on the next project. And of course as others have said the magnitude of the changes matters.

 

[rjc]

G71 was a thorough redesign of G70 for a brand new process (90nm) in a much smaller core, was on the market as a high-end for a mere 6 months, and precluded the process that was to be used on the "real" next generation, the G80, a very large chip by any standards.

What makes you think all of this can't happen again ?

I dont think 90nm was brand new for the G71. According to wikipedia the C51 on 90nm arrived in Nov '05 then the G72 and G72m in Jan 06 and the remaining chips were in march '06 - the G71, G73. So like a 5 month delay before they tackled the high end on a smaller process. So if they follow the same plan should be a integrated part followed by a low-mid range chip. As neither of these has shown yet for 40nm the GT212 cant appear before mid year if the same timetable is followed.

Am actually thinking they might be considering skipping the GT212 if GT300 is running close to schedule. If microsoft decide to introduce dx11 early could be kind of awkward trying to sell a new high-end that only supports dx10.

 

[Ailuros]

Simple reasoning would tell you that if the ballpark in terms of performance between GT212 and GT300 is large enough, there's no particular reason to skip anything.

212 appears hypothetically at timeframe X as single chip high end and slips down to the "performance" segment as soon as 300 hits the shelves. IMHLO the die size of the GT212 should be ~300mm2 despite its rumoured increase in clusters/SPs, meaning that they can easily sell such a chip as a performance and later on as a mainstream part without a problem.

 

[Ailuros]

double post