NVIDIA GF100 & Friends speculation
- Thread starter Arty
- Start date
Ninjaprime
Regular
I think thats only the GTX 480, I'm sure there will be a lot more GTX 470. Makes sense with all the problems of yields, ect they have been having.
Obviously they'll have more 470's than 480's and it's also convenient since 470 demand should be higher (due to lower MSRP). However I don't see how that is directly connected to yields. With the term yields my mind goes to a persentage of operational cores per wafer, ie the amount of cores that don't have any process related errors. Once you have a persentage there for operational dies the IHV will take those and check them if there are any faulty units and after that the frequency tolerance of the chip.
Given that TSMC itself assures that 40G yields are now roughly on 65nm level, if one knows where the GT200@65nm were it would be easy to make an estimate. For the frequencies it would be trickier to estmate at this stage anything since no one seems to know yet where those will be for the top dog. Obviously the higher the final 480 frequency will be, the smaller their final amount and there you actually have a persentage of chips with X frequency out of a A% persentage of operational dies.
I think its quite sad what nvidia is doing. Like one month before launch and probably with cards in production ( 
) and no fix clock rates, card info . Nvidia AIB partners needs to be silent like dead. GTX 480 and 470 fan facts realeased on facebook 
. Its like throwing bones to a hungry dog. Anyone seen a company of that size comunicating trough facebook ? Its almost embarassing
) and no fix clock rates, card info . Nvidia AIB partners needs to be silent like dead. GTX 480 and 470 fan facts realeased on facebook . Its like throwing bones to a hungry dog. Anyone seen a company of that size comunicating trough facebook ? Its almost embarassing Well im sure they could manage an HD 5890, aka ridiculous edition. However there are several important considerations.
1. Power/board.
They'd need a 6 pin + 8 pin PCI-E to faciliate the 230+ watt typical design power. This would also require beefed up VRMs. Luckily they can share the same board as the Eyefinity card which also uses 6 + 8 pin power. Ridiculous editions need ridiculous quantities of RAM as well, so probably 16 chips of 1Gb GDDR5.
2. Cooling.
Since the HD 5870 is kind of on the edge of its maximum acceptable cooling limits in terms of noise. So for most people they'd probably need vapour chamber cooling, which is luckily the forte of their referrence board maker Saphire. In addition to this, a back plate for cooling ram would not go amiss here.
3. Price.
As this is yet another board with a fully equiped RV870 die it would probably have to be at least $80 more expensive than the HD 5870. So therefore it fits into an interesting and small niche of people. Faster than the HD 5870 but slower than the 5970 except when crossfire doesn't work and then its quite a bit faster. They don't have to sell more than enough to prove that its a viable SKU, a Halo product really.
So what can they do with it? Well they'd probably be able to slap on a 3 Teraflop badge if they tried hard enough. Noone would buy it, but I guess the idea is the dream of buying it would/could influence people to get a lesser SKU off them instead.
Lets se a time frame. Lets say April - May for the 5890. That gives them 6+ months since the 5870 came out. They could have done a number of things . 1) a respin for better yields and clocks 2) they could have created a new chip with slightly more shaders and a slightly higher core clock .
Also for many a 1ghz 5870 with 2 gigs of ram would be ideal for eyeinfinty. Increasing the resolution to such great amounts and then adding fsaa will surely push a 1gig card to its limits.
Well again newer binned chips may perform better and wouldn't require such a cooler. Also a refresh with more shaders may not require greater cooling than the 5870.
Of course even such a cooler as your claiming above may be par for the course when considering the 470,480 and 5890
Well
1) The existing product line can shift in price. The 5870 can drop to $350 or under msrp. The 5850 can stay where it is really as it will have a large price gap still . ITs msrp is $259 so it will be slightly less than $100 price diffrence between the cards . A 5890 can fit into the $450 price bracket. Its still cheaper than a 5970 but will require less power , will surely be cooler and require a less powerfull powersupply and what not.
2) Performance wise it will be under the 5970 card but for hardcore gamers tehy may rather go with two 5890s for even better performance esp if they are 2 gig cards. It will surely scale better at higher resolutions than the 5970 when ram comes into play.
It really depends on the work involved. Simply binning 5870s that clock to 1ghz shouldn't require much effort.
I would really think it depends on how fast the nvidia parts are and how hard it will be for ati to match it. If they can't match it ehy will simply keep the line up as is and compete with price till their next gpu comes out. IF they can match it they will
Personaly I just want prices to drop as I've said.
ITs such a general statement though. Whats on 65nm levels ? Tiny 5450 size chips or massive fermi chips ? It can mean anything .
Aside from that these may simply be the amount of chips nvidia will have ready for launch. Even if yields are better now it will still take awhile for those chips to make it into the hands of nvidia and then its partners and then store shelves or internet shops.
Of course mabye this is just where fermi is . Perhaps they have very limited supply even with improved yields and will battle with this problem for a long time to come. It may be why they named it the 400 series instead of 300 series. Perhaps we will see 40nm 8800s back on the scene this time instead of renaming it the 9800 or gt 250 they will name it the 350 or what not and use that as a high volume chip to compete with ati's high volume chips till they can fix the problems with fermi and fermi will stay $500 graphics cards .
Will be interesting to see. This is reminding me so much of the 5800ultra launch though
The GTX280 was a 583mm2@65nm die last time I checked. Obviously the higher the complexity of the chip the lower the yields, but if you carefully re-read my post you'll see that I said that if someone knows where the GT200 yields ranged in 65nm he could make an estimate.
Just stating the obvious which is common ground for all IHVs. How's that connected to yields in any case?
They could run into shortages for the top dog if the frequency is set too high, but as I said we don't know the frequencies yet and it's still not directly related to yields.
What on God's green earth has the naming scheme to do with all that?
The decision to name their DX11 line GT4xx makes sense since the GT21x@40nm variants received already the GT3x0 description and the latter will obviously exist for some time until they can get replace in the future with low end members of the GT4x family. In any other case both the GT212 and GT214 have been cancelled afaik, so there doesn't seem to be anything in the channel for 40nm to cover for mainstream or performance.
One step at a time; let's see how the GF100 based SKUs will look like first and then we can see how the lower end parts might look like.
You don't seem to have a full picture then what kind of a FLOP NV30 was from many aspects.
Don't you first determine at which frequencies you get acceptable yields and not the other way around - i.e at which (potentially horrible) yield you get acceptable frequencies? The only scenario in which that would be remotely feasible is if the GTX 480 is close to the performance of the also non-existent HD 5970.
Groo The Wanderer
Regular
The first run on 65nm was yielding ~62.5% for both the 280 and 260 combined. It never really went up all that much, at least never hit the numbers NV wanted.
-Charlie
A respin could fix bugs which prevent them from clocking higher and it could improve yield. I doubt they would add extra stream processors as that would compicate their lineup, see RV790. However with yields being poor at the start, there is still the small possibility that every Cypress is infact a salvage part. Performance per watt is most important for Hemlock, overall performance per die is more important for a higher end refresh part. Im not sure if a refresh would significantly effect the former, however it may give them more room on the latter.
Agreed.
I don't doubt that they will try. Having both single die and dual dice cards with the 'performance crown' would be ideal for them. The dual dice Hemlock will probably retain its place, this leaves the single die Cypress model as a contestant. It'll be hard, and interpretting Richard Huddy's comments could imply almost anything. Since Evergreen was essentially done almost a year ago if you consider the Juniper samples shown, they could have a replacement out earlier this year and they may suffice with trading blows in the meantime, or even having a slower card in the shorter term.
FrameBuffer
Banned
Just a question regarding the bolded part.. what was so wrong (complicated) about the RV790 ? I only ask (being biased and all as a 4890 owner) since IMO it's probally one of the best "refresh" parts ATI has cobbled together (next to the 1900 v 1800). It allowed ATI to address some issues while the process matured some and still is very very competitive (particularly in price/performace). maybe I'm seeing it as consumer rather than looking at it from a manu's side.
Theres nothing wrong as a consumer, but from a producers perspective.
They currently have 3 enthusiast grade chips, the Redwood with 400SP, Juniper with 800 and Cypress with 1600 and 4 levels of parts. Those 3 chips yield 7 SKUs currently. However if they produce a still bigger chip they would have 4 enthusiast chips, and they would have 9 SKUs. However they couldn't use a bigger chip on Hemlock which would require them to still produce Cypress and it would confuse the pricing structure. They didn't produce at all that many RV790 chips and yet they still had to deal with semi defective parts with the HD 4860 which recieved very little marketing support from AIB partners and was often a poor deal relative to the HD 4870 and 4850.
That's too much to go by...
I thought 3rd-4th week of February was the most recent expected date? Could be wrong.
Edit: But I just read over at Hardocp that it's rumored end of March due to a respin. That sounds more likely to me as leaks would seem to be springing aplenty if we were in for a late Feb release, instead it's relative dead silence.
Edit: But I just read over at Hardocp that it's rumored end of March due to a respin. That sounds more likely to me as leaks would seem to be springing aplenty if we were in for a late Feb release, instead it's relative dead silence.
Last edited by a moderator:
Ninjaprime
Regular
Well, the way they keep pointing to how they said first quarter and pointing out the first quarter doesn't end 'till March 31st, I'd imagine they'll just ever bit of that time and not launch until last week of March, maybe even the last day of March.
I can't even remember when someone suggested february, it's been "march at earliest with no real availability" as long as I can remember
Similar threads
- Deleted member 2197
- Architecture and Products
