VR-Zone: G96, G94 and G92 @ 55nm

[mczak]

So, 9500GT should be compared with 8600GTS?

http://www.hothardware.com/Articles/NVIDIA_GeForce_9500_GT_Mainstream_GPU/?page=6

Hmm, ok. So it is faster than G84, after all (if you look at the 9500GT OC vs 8600GTS, they've got the same memory clock, the 9500GT OC has a 4% shader clock advantage and the 8600GTS a 11% core clock advantage, and the 9500GT OC is usually 5-10% faster). I guess the better z/color compression really makes a difference (maybe more than PCIE 2.0).
That's nice, and enough (for the stock model) to beat HD3650 (its main competitor for now) in every benchmark. The crazy overclock is no surprise, there's no reason why it shouldn't reach at least the level of a G94/9600GT (though possibly the ram of retail cards may not overclock that well...). Power consumption OTOH is slightly disappointing, I'd have expected it to do better than a 8600GT, but it actually seems to use slightly more power.
All in all, nothing wrong with it (if the price is right), though I still think this product should have appeared at least half a year ago...

 

[AnarchX]

G94B:
http://diy.pconline.com.cn/graphics/study_gra/0808/1385626_5.html
- 13.57mm * 14.18mm = 192.4mm²

But they measured G94A with around 255mm².

 

[Arun]

Uhhhh...
How the hell is it possible to reconcile these two links:
http://diy.pconline.com.cn/graphics/study_gra/0808/1385626_5.html
http://diy.pconline.com.cn/graphics/reviews/0801/1205247_2.html
Or, even better:
http://www.pconline.com.cn/images/h...nm96GT_0216.jpg&namecode=diy&subnamecode=home
http://www.pconline.com.cn/images/h...9600GT_0512.jpg&namecode=diy&subnamecode=home

Seems to me they don't have a very consistent way to measure. I think just looking at the way they measured the G94-A1 this time around, it's very clear they are overestimating its size, while they weren't the first time. Looking at say their G84 shot, they're probably also overestimating the size by a few millimeters there. Sigh, how fricking hard is it to get this right? *shrugs*

Anyway, seems very likely it's a straight 19% shrink (as 55nm should be), or something much nearer that than what pconline is implying here. BTW, as for defective G94/G94b, NV seems to be selling them as 192-bit/3C SKUs under the name of "9600 GS" to HP. Not a bad plan from a marketing/financial perspective, but I wouldn't really want to be the one buying that thing either...

 

[Jawed]

I was just waiting for you to throw a wobbly

Their technique might be screwed up by the "grout" around the die. Dunno really. Flatbed scanner pix at a known resolution seem like the best way of measuring.

Jawed

 

[Davros]

what process would they have to go to to get 2 8800gt cores into a chip no bigger than the current 8800gts (g80) ?

 

[Blazkowicz]

isn't that GT200b? if the shrink does 19% on that chip too, that gives 576/1.19, 484mm², precisely the size of G80! not counting some removal of redudancy if they do that.

though the current 8800GTS is G92, not G80, here a GT200 at 40nm would seem to fit.

 

[AnarchX]

Seems G96 gone C revision:

http://www.pcgameshardware.de/aid,657420/Test/Benchmark/Geforce_9400_GT_von_Zotac_im_PCGH-Test/

 

[Davros]

I suppose you all know about the 9400

"Slowly but surely, Nvidia continues to replace its old low-end graphics cards with new models in the GeForce 9 series. Today company quietly launched the GeForce 9400 GT for the value segment of the 3D graphics market. Based on NVIDIA's G96 GPU, the chip ships with 16 stream processors clocked at 1.4GHz, while the graphics core runs at 550MHz. NVIDIA's reference specifications also call for a minimum of 512MB of DDR2 memory clocked at 400MHz, with its 128-bit memory interface providing up to 12.8GB/sec of peak bandwidth, while texture fill rate stands at a paltry 4.4 billion texels/second (about a tenth of the GeForce 9800 GTX's fill rate). The firm claims its card "more than doubles the performance of its predecessor, the 8400GS, thanks in part to 65 nm process technology and a 128-bit memory bus.

Board partners are free to go beyond these specs however, already we've seen a handful of partners clocking their boards at higher memory speeds than 400MHz, while most seem to be sticking with the stock speeds for the graphics core and shaders.The GPU has a 50W TDP with a max GPU temp of 105 degrees Celsius and ships with a single-slot cooler and 6.6" long PCB. GeForce 9400 GT cards will ship with two dual-link DVI display outputs with additional support for HDMI and digital audio via adapter. Cards are shipping now and should be hitting retail in the coming weeks. With its higher clocks and 16-shader architecture, the card officially replaces the 8500 GT in NVIDIA's lineup. Prices will start at $59."

 

[AnarchX]

High-clock G96B/C or G94B@192-Bit/48SPs?

edit: They say G94B 192--Bit&64SPs:
http://gpucafe.com/2008/08/nvidia-preparing-to-counter-attack-in-the-sub-150-segment/

 

[mczak]

High-clock G96B/C or G94B@192-Bit/48SPs?

edit: They say G94B 192--Bit&64SPs:
http://gpucafe.com/2008/08/nvidia-preparing-to-counter-attack-in-the-sub-150-segment/

This looks like trouble for nvidia to me, based on what we know on these cards.
9600GT vs 3870 - fair enough.
9550GT vs 4670 - based on these speculations, sounds ok too. However, the HD4670 should have a large advantage in manufacturing cost (I guess nvidia will try to save costs by using 384MB ram vs. the 4670 512MB, so it will tank in higher resolutions/AA against the 4670, but this should easily be outweighed by the simpler pcb and cheaper (smaller die) chip of the 4670).
9500GT vs 4650 - I can't see how the 9500GT would be really competitive. Now, the 4650 will be likely quite a bit limited by ram bandwidth, but still... At least nvidia shouldn't be at a disadvantage wrt manufacturing costs - more expensive (gddr3 vs. ddr2) ram, but cheaper chip.
9400GT vs 4450 - this looks awful. Not only is the 4450 placed above the 9400GT, but it should be cheaper to produce (9400GT is using G96, so comparatively large chip and complex pcb due to 128bit).

 

[Domell]

I think the best that NVIDIA could do against new ATI GPUs is push the 9550GT (G94B) with full 9600GT specs with only lower clocks (something like Radeon 9550 was underclocked 9600Pro). There will be great OC potential and 256-bit MC so NVIDIA could really win this part.

 

[Tchock]

ATI's partners can push RV730 easily too. The keyword here is simply "memory". The core itself should have no problem beating G94 given ample bandwidth.


I'm guessing partners will be pushing GDDR3 to ever-so-high records on midrange cards. And still profit more from OC editions. (GDDR3 is still cheap compared to the GPU IIRC)

RV740 is rumoured to have ~8800GT performance. Not sure whether it's extra compute power, or bandwidth liberation thanks to GDDR5.

BTW Komplett has the 4670 listed up there for 72 Euros, so I have the idea that nVidia's graph is totally off the charts.

 

[Domell]

ATI's partners can push RV730 easily too. The keyword here is simply "memory". The core itself should have no problem beating G94 given ample bandwidth.


I'm guessing partners will be pushing GDDR3 to ever-so-high records on midrange cards. And still profit more from OC editions. (GDDR3 is still cheap compared to the GPU IIRC)

RV740 is rumoured to have ~8800GT performance. Not sure whether it's extra compute power, or bandwidth liberation thanks to GDDR5.

BTW Komplett has the 4670 listed up there for 72 Euros, so I have the idea that nVidia's graph is totally off the charts.

Maybe true but how do you know what NVIDIA will offer in Q1-Q2/2009 when Rv740 is going to be released? Maybe NVIDIA will do something based on GT200 architecture to midrange GPUs in 40nm(i.e. GT210 or GT220 codenamed)?

 

[Tchock]

I'll be anticipating how much they can squeeze into ~150-200mm^2, disregarding memory interface width because GDDR5 is a possibility.

I don't think they'll be able to exceed the 9600GSO/8800GS specs at the designated diesize. They could do G80-esque specs, but I think it'll still be a stretch. Probably have to dump DP ALUs too.

 

[bearmoo]

RV730 review out ...

http://publish.it168.com/2008/0901/20080901043801.shtml

 

[bearmoo]

nice die shot

 

[ninelven]

It seems impressive from a die size point of view, but as far as perf/$ not so much. It loses pretty handily to a 3850 which loses to a 9600GT which can easily be found for less than $120 right now.

 

[Tchock]

Grain of salt. Fast.

After crosschecking IT168, it seems that they were the ones who first unveiled 4850/4870 performance too.

(More like, what a load of bu**. )

So grab your salt shakers slowly in regards to RV730 performance... the target resolution's 1440 and below anyway, which is obvious for the pricepoint.

 

[mczak]

It seems impressive from a die size point of view, but as far as perf/$ not so much. It loses pretty handily to a 3850 which loses to a 9600GT which can easily be found for less than $120 right now.

Keep in mind the 4650 should be similar priced to the 9500GT - and there's no contest there really. I'd agree the 9600GT is quite close in price and might be a better value if you look at performance/$, but still it doesn't compete in the same price bracket (there is, after all, even the 9550GT between the 9500GT and 9600GT too).

 

[ninelven]

Keep in mind the 4650 should be similar priced to the 9500GT

I was just going by the AMD slide....