NVIDIA GF100 & Friends speculation

[mczak]

If I'm not mistaken, Cypress has 20 80-wide SIMDs, so 80/1600 = 5%, not 10.

If I'm not mistaken, it is not possible to disable only one SIMD, need to have the same number of simds in both rasterizers. I could be wrong though...

As for GF104 following in GF100's footsteps, well… for now there's no indication of it being released with the full chip enabled, which doesn't exactly bode well.

This is true, but you'd think with a) smaller chip, b) learned from mistakes from GF100, and c) 40nm manufacturing still getting better things would improve?

 

[Alexko]

If I'm not mistaken, it is not possible to disable only one SIMD, need to have the same number of simds in both rasterizers. I could be wrong though...

I have no idea, I just assumed it was possible.

This is true, but you'd think with a) smaller chip, b) learned from mistakes from GF100, and c) 40nm manufacturing still getting better things would improve?

I'm not sure they really had time to feed the experience they gained from designing GF100 back into GF104. After all, there's only a 3 month gap between their respective release dates.

 

[smw]

If there are a lot of small errors crippling in larger dies (>300mm^2) it may be viable for top tier (i.e. Cypress and GF100) to launch full scale chips at premium prices even on small quantities (RV870 already has it's full chip enabled in the HD5870 and a recent article suggests that there might be a full on GF100) I don't really think that such a thing would be reasonable for a chip that is launching in the $200-250 price bracket if it would bring them less margins because of additional testing required to make sure that the full chip is indeed functioning properly. My reasoning comes from 1) a post Dave Baumann made regarding the high power consumption of the HD5830 that even tough the clusters were disabled for use they still consumed the power and that the consumption was more a function of the clocks and 2) there were reports of GF100 GPUs with all clusters enabled around it's launch. So if 1) also hold true for NVidia and a full chip wouldn't have increased the power consumption and it still didn't make business sense for them to release a GTX485 (or some such) then I don't know why you think that they will release such a beast with their value proposition in the face of GF104. I'm not saying that they can't or won't do it (i.e. in the premium cards from EVGA or other company where the premium price would also make sense or when there are enough good chips coming out that can easily meet the testing) but just that if the 40nm process is still rather shitty it may be cheaper for them to keep it as it is now.
Just my 2 cents
PS: sorry about the rather lengthy non-technical post

 

[Picao84]

What about some paragraphs?

 

[Dave Baumann]

I think more fuss should be made over redundancy in HD5850. Why are cores being lost when RV770's salvage model seemed fine.

Redundancy is not free, it costs area. You have to consider over the course of an entire product line what is more costly. Would rather have every Cypress as 18 SIMD's?

 

[smw]

What about some paragraphs?

Well it's about a single topic and a single line of thought so it doesn't feel natural to have many paragraphs but if it is unreadable I don't mind for the mods to chop it into pieces as they see fit.

 

[Jawed]

Redundancy is not free, it costs area. You have to consider over the course of an entire product line what is more costly. Would rather have every Cypress as 18 SIMD's?

So, there's no 17th ALU per SIMD in Cypress?

 

[Jawed]

I really don't know why people are surprised by the disabling going on or expect full chips from NVidia as recently they always left at least one cluster for redundancy - both GTX280 and GTX285 also had one cluster disabled

I stopped reading there. Go check your facts.

Jawed

 

[smw]

I stopped reading there. Go check your facts.

Jawed

To be honest I'm kind of tipsy right now so I might have written some nonsense but the rest still stands on it's own. As for the text quoted by you I seem to remember there were rumors that in the GTX280 and GTX285 there were 256 ccs but only 240 were enabled. May be they were just that - rumors. What I don't remember is if they were proven wrong or right.

EDIT: whoops! I just checked and at least from the diagram at xbit there isn't a place for more shaders to be placed so yeah - brain fart there. I still believe that the rest is valid though - if it doesn't make sense to them business wise to enable the cores in GF104 I don't think they'll do it just to prove they can.

EDIT2: just to delve further into the business side of things - at least on the marketing slides it was shown that they plan on releasing the cards at $200 for the 768MB variant and $230 for the 1GB one. So a full spec GF104 chip could easily obsolete the GTX465 which is $280 at newegg right now. I think that as long as there are enough "not good enough" chips to support the 465 line they will not release such a thing as a full spec GF104 and then they can pull it off the market as to not have to disable working cards in order to supply that same line (4830 style). When such a moment comes they can replace the 465 GPU with a full fledged GF104 instead of a crippled GF100 (and call it a GTX465 rev 2 for example) but as it is now I think it is more important for them to sell the expensive GF100 dice rather than throw them away

 

[TKK]

(...) And seeing these GTX460 vs GTX465 numbers, it looks like GF104 is just somewhat "slower per theoretical ALU throughput" to me, though I wouldn't know why. I guess it has less SFUs (per normal ALUs), but I'm not convinced that's important here. Maybe the cache changes really do make a performance difference in practice.

Personally, I believe it's mostly the combination of
- less pixel fillrate
- less memory
- less mem bandwidth.

All of these Gigabyte tests were done at 1920x1200 with 8x(!)AA, for example I'm rather sure the bad result in Unigine was not just because of the halved geometry performance, but also because 768MB simply aren't enough with these settings. Tesselation is a rather memory-hungry feature itself, combine this with a high resolution and high level of AA, and exceeding 768MB should be easy.
I expect the 1GB version to be quite a bit faster at high resolutions with 8xAA.

 

[onethreehill]

GTX 460 768MB Pictures
http://vga.it168.com/a2010/0709/1075/000001075571.shtml

 

[Florin]

GTX 460 768MB Pictures
http://vga.it168.com/a2010/0709/1075/000001075571.shtml

Yuck, it's ati-red.

 

[fbuffer]

GTX 460 768MB Pictures
http://vga.it168.com/a2010/0709/1075/000001075571.shtml

isn't dual 6 pins (up to 225W) a bit overkill for a "mainstream" part ? Honestly I was expecting 150W at most putting in place at around the 5770 however this puts it around 5850 range.

 

[3dcgi]

Personally, I believe it's mostly the combination of
- less pixel fillrate
- less memory
- less mem bandwidth.

All of these Gigabyte tests were done at 1920x1200 with 8x(!)AA, for example I'm rather sure the bad result in Unigine was not just because of the halved geometry performance, but also because 768MB simply aren't enough with these settings. Tesselation is a rather memory-hungry feature itself, combine this with a high resolution and high level of AA, and exceeding 768MB should be easy.
I expect the 1GB version to be quite a bit faster at high resolutions with 8xAA.

Tessellation isn't memory hungry. It's a form of compression.

 

[tannat]

As for GF104 following in GF100's footsteps, well… for now there's no indication of it being released with the full chip enabled, which doesn't exactly bode well.

Is this due to yield issuses?

They may just hold the chips until GF100 GTX465/GTX470 have cleared the market.

 

[psolord]

isn't dual 6 pins (up to 225W) a bit overkill for a "mainstream" part ? Honestly I was expecting 150W at most putting in place at around the 5770 however this puts it around 5850 range.

They say that the GTX 460 768MB will be 150W and the GTX 460 1GB will be 160W. 160W is already past the one 6pin+PCIe slot wattage. Not that it wouldn't work anyway, but still specifications are to be followed.

Moreover, they officially slided the OCability of these babys, so power consumption will fly, once you overclock it and overvolt it to hell and back.

All I want from these cards is to have a very low idle power consumption. I mean around 20W and no more. Hope they fixed this.

 

[neliz]

They say that the GTX 460 768MB will be 150W and the GTX 460 1GB will be 160W. 160W is already past the one 6pin+PCIe slot wattage. Not that it wouldn't work anyway, but still specifications are to be followed.

Moreover, they officially slided the OCability of these babys, so power consumption will fly, once you overclock it and overvolt it to hell and back.

All I want from these cards is to have a very low idle power consumption. I mean around 20W and no more. Hope they fixed this.

http://www.fudzilla.com/graphics/graphics/graphics/nvidias-gtx-460-to-be-great-overclocker

It's all in there, ofcourse, the factory OC models will have a bit higher power consumption.. but when everybody is talking about good OC/Performance and nobody is talking about power consumption, what would be your bet? To put the GTX460/768 in perspective, it's "nVidia TDP" is the same as the HD5850 TDP (150nvidiWatt, versus 151Watt).

I personally think we'll see factory OC models near 800Mhz too but I guess that will be after the first launch wave, with things like dual fan cooling solutions.

GF104 can't be compared to Juniper, simply because it's in a different price class. There's like a $50/$60 gap there. GTS450/455 will be more in spec. with Juniper.

 

[onethreehill]

GTX 460s @newegg
http://www.newegg.com/Product/ProductList.aspx?Submit=ENE&N=100007709%20600030348%20600007323%20600062521&IsNodeId=1&name=GeForce%20GTX%20460%20%28Fermi%29

This is an interesting card
Palit GTX 460 SONIC PLATINUM 1GB 256-bit
Core Clock 800MHz
Shader Clock 1600MHz
Effective Memory Clock 4000MHz

 

[Cookie Monster]

Effective Memory Clock: 4000MHz

Interesting. Maybe they managed to fix some of issue surrounding the memory clock problem.

 

[CRoland]

Interesting. Maybe they managed to fix some of issue surrounding the memory clock problem.

That's the memory frequency of stock 5830/5850 on an overclocked 460 1GB...