NVIDIA's New Entry-Level GF108 GPU Pictured

[Valdez]

Originally Posted by claylomax

Is it like a £2 coin, because I can't figure out the size of the chip then.

The 100 Ft has a diameter of 23,4 mm.

[pantherx12]

Originally Posted by claylomax

Is it like a £2 coin, because I can't figure out the size of the chip then.

Compare it to the size of the chokes instead, they look fairly standard.

[vagxtr]

Originally Posted by a_ump

definitely. no company's ever put more than that on entry level graphics. the 9600GSO was mainstream at release so that doesn't count :P.

so its gonna have 128SP's yet perform worse than or around GT 240 level's. Nice, more products from nvidia that are less efficient per sp than the G200 core.

It can hardly perform worse than Geforce GT240 which had only 96SPs and that were those bulky 24SPs/cores from GT200 and only 8ROPs. GF100 might not shine in all it's glory ... it's ugly oversized product like first GT200 and bigger on much worse node than highly reliable 55nm (which was much easier and faster implemented)
So in fact we should expect good performance of this chip if it support 128 bit memory bus which it should cause it support budget ddr3 chips and even 3year old g84 aka 8600gt came out with faster gddr3-1600 ntm gddr3-2200 on 8600gts. It would probably overtake 320SP Redwood chips carrying gddr3-1600 in 5550/5570 case. Somehow i dont believe it will follow GT215 (GF GT240) 8ROPs only wrong philosophy, and feature like full 16ROPs just like GF106 and other mainstream predecessors G84/G92/G94 from the past, and with that be much closer in performance to HD5670 even with limited gddr3 128bit bus. But it will probably consume far more power than 35W (HD55X0) and more like 55-60W like GDDR5 equipped HD5670

[a_ump]

o no but it can lol. u know the GT 240 is only on par with the 9600GT half the time and the 96GT only has 64 SPU's, but the higher clock def helps along with the 16ROPS. Really could be anything. If it's 128SPU's are clocked high around 650 or higher, with 16ROP's(8800GTS/9800GTX config) then yea should be between the 9800GT. Would be kewl to see if nvidia ever come out with a card that has a high clock. Course...why couldn't it have say 12 ROP's?

I mean if this comes out with say an 875core and 17XXmhz shaders, would be a bomb card. But that's just my imagination.

[KainXS]

the difference is that the GF106 uses GDDR5, it might not seem like much of a difference, but it is

[a_ump]

....so do a lot of GT 240's, esp the review samples, and it with 96 shaders, performs dam near always lower than the 9600GT the majority of the time.

[url=http://www.techpowerup.com/reviews/MSI/GeForce_GT_240/30.html]GT 240 perf summary[/urll]
it always loses to the 9600GT, so who knows, really is up in the air how the GF106 will perform

[vagxtr]

Originally Posted by a_ump

....so do a lot of GT 240's, esp the review samples, and it with 96 shaders, performs dam near always lower than the 9600GT the majority of the time.

[url=http://www.techpowerup.com/reviews/MSI/GeForce_GT_240/30.html]GT 240 perf summary[/urll]
it always loses to the 9600GT, so who knows, really is up in the air how the GF106 will perform

Of course it performs worse when GT240 (g215) is damn 128-bit card with ONLY 8ROPs while 9600GT (g94) was fully featuring 256-bit card. It means a lot in gaming/3D when you use only half memory bandwidth.

In case of GF100 series they ALL USE GDDR5 (guestimation --- they wont do unnecessary redesign job just to put gddr3 onto lowest end card, but it could probably work with dirty cheap ddr3 (different bstd)). So GF108 if wouldnt be put onto budget 64-bit pcbs for budget market will probably sport GDDR5 128-bit memctrl which will give it on pair (or better?) memory bandwidth as g92(a/b)/g94 chips had (8800GT/GTS, 9600GSO, 9800GT/GTX/GX2/GTX+, GT230v2, GTS240/250, (Dual GTS250 by Zotac) for g92 and 9600GT/GSO512, GT230v1 for g94).
Pretty big envydia's product line with only two chip designs

GF106 which you mentioned will be part of lower mainstream GTS440/450 and it'll sport 192-bit bus and be ready to whoop GTX260-216 ass (by some announcements )

[Yukikaze]

Originally Posted by vagxtr

Of course it performs worse when GT240 (g215) is damn 128-bit card with ONLY 8ROPs while 9600GT (g94) was fully featuring 256-bit card. It means a lot in gaming/3D when you use only half memory bandwidth.

9600GT packs GDDR3 over a 256-bit bus, while the GT240 packs GDDR5 over a 128-bit card. While the ROP difference is noticeable, there is practically no difference in memory bandwidth (IIRC, I don't remember the exact clocks, but at identical base clocks, GDDR5 over 128-bit is identical to 256-bit GDDR3).

[vagxtr]

Originally Posted by Yukikaze

9600GT packs GDDR3 over a 256-bit bus, while the GT240 packs GDDR5 over a 128-bit card.

but at identical base clocks, GDDR5 over 128-bit is identical to 256-bit GDDR3).

Thank you for confirming what you just quote

I think it's pretty clear now, 26 month after ATi's hd4870 release, what gddr5 brought to the world over gddr3 and it's called memory bandwith (doubled per pin (aka.bit) in case gddr5 vs.gddr3 (if they work on the same base clock))

While the ROP difference is noticeable, there is practically no difference in memory bandwidth (IIRC, I don't remember the exact clocks, ...

There's in fact huuuge practical difference in the number of ROPs in envydia's case (hint every pair of ROPs in G80/GT200, and every quad-ROPs in GF100 generation were connected to 32-pin (bit) memory line)