NVIDIA GF100 & Friends speculation

[mczak]

Well if it made sense for Cypress (HD5830) it definitely makes sense for GF100.

Ok. Well in contrast to the lobotomized Cypress where half the rops also mysteriously seemingly halved the bandwidth available to the other rops, it should scale down much more gracefully.
They can cut out quite some more stuff if it only should be faster than HD5830, though if they want to sell it at HD5850 prices I guess not too many more clusters can be sacrificed. Maybe one or two if they can at least reach GTX470 clocks and even that would be quite a stretch in the competitiveness business imho...

 

[Picao84]

Ok. Well in contrast to the lobotomized Cypress where half the rops also mysteriously seemingly halved the bandwidth available to the other rops, it should scale down much more gracefully.
They can cut out quite some more stuff if it only should be faster than HD5830, though if they want to sell it at HD5850 prices I guess not too many more clusters can be sacrificed. Maybe one or two if they can at least reach GTX470 clocks and even that would be quite a stretch in the competitiveness business imho...

At least GTX470 clocks? Shouldnt it be way higher? At least the potential.. Of course nVIDIA doesnt want it to get too close to GTX470..

 

[Blazkowicz]

The dual 6-pin connectors lead me to believe it's another cut down GF100 die.

it's a GTX 470 shown for illustration purpose.

 

[trinibwoy]

it's a GTX 470 shown for illustration purpose.

Completely missed that, thx.

 

[Ailuros]

The dual 6-pin connectors lead me to believe it's another cut down GF100 die.

GF104 will most likely end up with 2*6pin connectors too

 

[Arty]

http://en.expreview.com/2010/04/08/geforce-gtx-460-to-arrive-in-early-june/7014.html

So a lower performing part priced at 5850, gotcha.

 

[mczak]

At least GTX470 clocks? Shouldnt it be way higher? At least the potential.. Of course nVIDIA doesnt want it to get too close to GTX470..

Well if that's another cut-down GF100 part to get rid of the "crap bin", I just don't see nvidia raising clocks to more than what GTX470 has.
Of course, you'd expect GF104 to have somewhat higher clock. I wouldn't expect miracles though at this point. High clocks didn't happen for gt200b nor any of the gt2xx derivatives (of course, gt215 is the only one where it even matters from those). So I'd consider anything slightly above GTX480 levels but with much better perf/power a success.

 

[KimB]

Well if that's another cut-down GF100 part to get rid of the "crap bin", I just don't see nvidia raising clocks to more than what GTX470 has.
Of course, you'd expect GF104 to have somewhat higher clock. I wouldn't expect miracles though at this point. High clocks didn't happen for gt200b nor any of the gt2xx derivatives (of course, gt215 is the only one where it even matters from those). So I'd consider anything slightly above GTX480 levels but with much better perf/power a success.

I don't think it'd have a different codename if it was just a different binning of the GF100.

 

[CarstenS]

GF104 will most likely end up with 2*6pin connectors too

So powerful in performance or so wasteful with energy?

 

[Ailuros]

So powerful in performance or so wasteful with energy?

Isn't there a chance those two factors go hand in hand for the underlying architecture? It would be definitely nice to see a hypothetical smaller core lying somewhere underneath of 470 performance (but not too far apart wherever there are no bandwidth bottlenecks) with a 1*6pin layout, but I'm not that optimistic.

 

[Mindfury]

GTX 460 performance leaked:

http://translate.google.com.hk/tran...ticle/2010/0428/A20100428645765.shtml&prev=_t

Performance 3Dmark Vantage's Extreme settings (1920x1200 4x AA), GTX460 scores roughly x5700, x6100 compared to 5830's slower on the level of about 7%. Game performance, Crysis Warhead 1920x1200 VHQ performance as 23FPS, similar to the basic and 5830.

 

[Mindfury]

GTX 460 是 GF104 手下败将。
Edison 发表于 2010-4-28 16:03

GF104 would be stronger than GTX 460 at game performance.

 

[trinibwoy]

Asked the gf and she translated it to "GTX460 is a loser compared to GF104"

 

[wishiknew]

I translate more like GF104 already defeated the GTX460.

 

[Tchock]

For the lulz:
http://www.fudzilla.com/content/view/18608/1/

So according to Fudo GF104 has no cache

But even if we take his comments with just a semblance of truth, the L2 is gone?

should be seen as the new Geforce 8800GT

the rumored die size would be just... that.

 

[fellix]

No R/W cache?!

**************

BTW, I've managed to obtain some numbers from an overclocked GTX480 in the good old ArchMark:


http://imageshack.us

The GF100 entry is in the last column, next to Cypress.

Note the quite high values for the depth tests -- caching?!

 

[KimB]

No R/W cache?!

**************

BTW, I've managed to obtain some numbers from an overclocked GTX480 in the good old ArchMark:


Uploaded with ImageShack.us

The GF100 entry is in the last column, next to Cypress.

Note the quite high values for the depth tests -- caching?!

I'm thinking compression.

I do find it interesting that it seems that Cypress is really good at certain tests you'd expect to occur more often in gaming, but has huge hits in some other places, while nVidia (mostly) does about the same across the board for similar loads. Like how nVidia is about the same with RGB and BGR, but ATI is much, much worse with BGR.

 

[mczak]

GTX 460 performance leaked:

http://translate.google.com.hk/tran...ticle/2010/0428/A20100428645765.shtml&prev=_t

Interesting that 5 (not the previously assumed 4) SMs are disabled. That's a lot - not unlike the HD5830 though which disables 6 out of 20 clusters. Shouldn't have the "ROP problem" the HD5830 has however, but according to that link it will cost the same as a HD5850 which clearly isn't competitive.

 

[mczak]

GF104 would be stronger than GTX 460 at game performance.

Asked the gf and she translated it to "GTX460 is a loser compared to GF104"

I translate more like GF104 already defeated the GTX460.

I don't see much difference here, honestly .

Anyway, it's not hard to see that gf104 could potentially beat (though I remain sceptical) a GTX460. Granted GTX460 has 11 instead of 8 SMs, but presumably GF104 could make about half that deficit up with higher clocks. And with way more tmus, and (presumably) higher ROP and memory clock (with same rop count and memory interface width as GTX460), that could be enough to beat a GTX460. With the reported die size though, unless this could compete with a HD5850 (which I seriously doubt), it doesn't seem to pose much of a threat to AMD, but if it doesn't have the fabrication issues of GF100, much more viable products should see the light of the day.

 

[wishiknew]

I got another translation. GTX460 lost by the hands of GF104. GTX460 bested at the hands of GF104.

Yes they all do mean the same thing but I need practice.