NVIDIA GF100 & Friends speculation

[Kef]

I don't believe so. They labeld the dx 11 one as dx 11 and it has lower performance for ati than the same res non dx 11.

They just screwed up something.

Maybe, I don't know. But they are _not_ running the same code-path.

 

[3dilettante]

I think it's more a "virtual" structure than a real one just invented to have something resembling somehow the old TPCs . It would make more sense (just think of the load balancing issue between GPCs) if this structure simply does not exist in the way it is drawn in the slides.

If you look at the some of the high-level die shots, there does look to be an area of logic that looks like it was replicated 4 ways, and identically so.

The SMs are arranged in groups of 4. The vertical dividing line in the middle of each group of four touches a smallish box of logic in that center band that is identical at each place neighboring the groups of four. There is a high level of symmetry in the center region, but there are some deviations in some corners that show it is not entirely symmetrical. That bit of logic is symmetrical despite the other variations.

It's at the very least not inconsistent with some kind of 4-way subdivision of resources in the full chip. Just what is divided 4-ways is something I don't really know, just that something visually appears to be split in that fashion.

 

[corduroygt]

While performance seems competitive, power consumption and heat are very disappointing. I'd like to see how the smaller versions compare against the 5770, which I consider to be in the sweet spot for price/performance/heat.

 

[psychoticdream]

Yeah, but they have two graphs of DIRT2 at 26x16x4aa and they show a 24% drop on the 5870 from the DX? to the DX11 graph. Very odd. Maybe they had vsync on

apparently it was the demo

on the 480 it defaults to dx9 which might account for the HUGE jump in frames

 

[Creig]

Seems to me that the 5870 is framebuffer limited in some of these.

I was thinking the same thing. I'm curious to see the 480 benchmarked alongside a 1GB 5870 and a 2GB 5870. Then we can get an accurate performance picture and see how much a 1GB framebuffer is limiting performance at 2560x1600.

 

[Mize]

apparently it was the demo

on the 480 it defaults to dx9 which might account for the HUGE jump in frames

No, that doesn't explain it because they have two separate graphs for 25x16x4aa ultra and they have different numbers for HD5 on those two graphs. They had to run the HD5s two different ways.

 

[eastmen]

http://img.hexus.net/v2/graphics_cards/nvidia/Fermi/N174/OCb.jpg

800mhz on air with a 101c temp fan speed at 80

 

[Kef]

No, that doesn't explain it because they have two separate graphs for 25x16x4aa ultra and they have different numbers for HD5 on those two graphs. They had to run the HD5s two different ways.

Real reviews coming up. Why bother with propaganda?

 

[Mintmaster]

Looking at the Hexus numbers, here's what I see:

When the 5870 is well below 60 fps and the resolution is still playable, the GTX 480 is notably faster in FarCry 2 and HAWX at 2560x1600 8xAA. That's it. And it consumes 30W more when idling and 140W(!) more under load.

That's not a particularly compelling argument for buying one. I guess you can extrapolate the useless wins to future games, but it's not a very consistent trend.

 

[rpg.314]

Real reviews coming up. Why bother with propaganda?

Coz the real deal is ~4 hours out and this is B3D.

 

[psychoticdream]

I was thinking the same thing. I'm curious to see the 480 benchmarked alongside a 1GB 5870 and a 2GB 5870. Then we can get an accurate performance picture and see how much a 1GB framebuffer is limiting performance at 2560x1600.

if frame buffer IS a problem that would explain why nvidia went with 1.5 of memory instead of the 1gb expected at first

 

[Kef]

Coz the real deal is ~4 hours out and this is B3D.

Sweet, but not even semi-accurate.

 

[rpg.314]

if frame buffer IS a problem that would explain why nvidia went with 1.5 of memory instead of the 1gb expected at first

Actually, I think 512 bit bus was expected before the GTC reveal. So 384 bit would be a step down.

 

[Razor1]

Singing a bit too early aren't we?

not really there is much more to it then what Sontin just stated

 

[Kef]

not really there is much more to it then what Sontin just stated

Please share.

 

[rpg.314]

Sweet, but not even semi-accurate.

Which bit? The info-strip-time or the domain name of this forum? :smile:

 

[Kef]

Which bit? The info-strip-time or the domain name of this forum? :smile:

All the bits. I don't care if they are 256, 320 or 384 bits

 

[Razor1]

Please share.

you will in about 3.5 hours

 

[Kaotik]

Gah, this is annoying, I'm tempted not to go to bar to get wasted, and rather stay home and get wasted for proper reviews, but I suppose I'll go to bar anyway, pick up a girl and bring her home and when we're here I'll just tell her to sit tight, I need to read through some reviews first

edit: kitch9 @ FM, if you happen to read this, I must say that it's quite annoying not to be able to post there atm, regarding other things than GF100 too, but ye, the DiRT2 DX9 numbers are definately wrong

 

[Kef]

you will in about 3.5 hours

I doubt I will be impressed. But if I'm wrong, you'll be the first one to know.