Fermi NDA ends today

[Jokester]

Just so we can quote you guys for future reference, you are publicly stating you think the new single card high end from nvidia will be faster than the 295

You can quote me as I was just repeating what one of the reviewers from Bjorn3d (a respectable site I think?) posted before it was removed. At the end of the day, Fermi has close to double the units that the 285 has and each unit isn't just a copy of the last gen, put completely reworked to improve efficiency in the same way that G80 did over the 7 series generation (processing not thermal ).

 

[Duran]

Driver updates have pushed the 5870 passed the junk that is the 295 in a number of games, ATI got the job done.

still slower overall. the gtx 295 is a great card that was the top performer for almost a year, overlocks well with low noise output and sensible temperatures. what more could u ask from a dual gpu solution? compare that to the 4870x2 that sounds like a vacuum cleaner, hotter than the sun with no overlocking capabilities... thats what i would call a piece of junk. you get what you pay for after all

 

[Rroff]

Aside from a small number of benchmarks and a couple of games 200 series SLI setups are still faster than the 5870 in the majority of games.

 

[LoadsaMoney]

Birthdays 24th March, if these are out by then ill treat myself

 

[drunkenmaster]

it wasnt long ago that ati owners kept claiming that a single card from a new generation can never beat the top dual gpu from the previous generation. of course just to justify the poor performance of the 5870

Incorrect, a single new gpu is highly unlikely and to date has never beaten the previous generation top gpu x2, which is very different to a dual gpu version. In this case a 295GTX is quite significantly slower than 2x280's or 2x285's in SLI. The odd game they'll be very close, but most there is a very noticeable difference.

Likewise the 285GTX was a refresh with higher clocks, and so was the 4890, they weren't the "architecture, they were architecture + steroids".

THe 5870 is what a good 70-80% faster than the 4870. Like the Fermi would be expected to be closing in on twice the speed of a 280GTX(and won't make it), like Fermi will almost without question NOT beat 2x285gtx's in SLI but could well beat, and probably will beat a 295GTX, but in worst case scenario theres a good 20-30% difference between 285gtx sli and 295gtx.

A 8800gtx spanked a dual gpu mobile version previous gen dual GPU card, but was slower in pretty much 4/5 games than x1950xt's in Crossfire.

This trend has been around for a VERY long time, and refreshes are rarely counted as the card they are aiming to double on performance, because with refreshes you're normally looking at mature process's or in the case of a 285GTX, a new process entirely, versus the speeds capable on a new process. New cards, new process, new refresh on mature process's with often slightly tweaked cores. I would hazard a guess that a theoretical 5890 would be a good 15-20% faster than a 5870, which would also help it get to the 70-80% faster than a 4890 bracket.

You can simply forget about a 20-30% here and there and pretend claims were made that are untrue, but they aren't. This gen, partially due to the process, partially due to the architecture and shift to DX11 you're getting two things, more than 50% larger in size than previous gen on a low yield process so they aren't price competitive compared to last gen. Also with added architecture you'll find they should scale better in DX11 than they would in DX10.

What if in the same game the DX11 5870 is 110% faster than a 4870 running the game in DX10.


IF and its a HUGE if those video's of a 285gtx vs a Fermi in Farcry 2 were real, then you'd be looking at a not completely unexpected, roughly 66% increase in speed and NO WHERE NEAR a 100% increase.

 

[drunkenmaster]

Aside from a small number of benchmarks and a couple of games 200 series SLI setups are still faster than the 5870 in the majority of games.

THe Fermi will not beat 200 series in SLI, (295gtx and 260 192 versions aside), like Nvidia have also never once beaten the previous Gen in crossfire with their top card, so whats your point exactly, 2x£300 cards a year ago are about the same speed as a single £300 card this year? Which has been the pattern for the industry for as long as SLI/Crossfire have been around?

 

[Marine Iguana]

Cant wait to see if anything interesting pops up not gonna go ATI anytime in the future.

 

[Owenb]

You can quote me as I was just repeating what one of the reviewers from Bjorn3d (a respectable site I think?) posted before it was removed. At the end of the day, Fermi has close to double the units that the 285 has and each unit isn't just a copy of the last gen, put completely reworked to improve efficiency in the same way that G80 did over the 7 series generation (processing not thermal ).

The G80 was a huge step because it was the first unified shader architecture, don't expect that sorta jump with Fermi.
It HAS to be faster than a GTX295, otherwise it's in the same ballpark as a bandwidth limited 5870.

 

[Rroff]

I think you missed my point entirely.

 

[Lightnix]

You can quote me as I was just repeating what one of the reviewers from Bjorn3d (a respectable site I think?) posted before it was removed. At the end of the day, Fermi has close to double the units that the 285 has and each unit isn't just a copy of the last gen, put completely reworked to improve efficiency in the same way that G80 did over the 7 series generation (processing not thermal ).

Close to, you say? So presumably 448 shaders. Okay. Also, I'm not sure I'd call Bjorn 3D a 'respectable' site. In the past they've been big fans of non-zeroed graphs to make Nvidia look good, but they're probably quite reliable on the early press-release front.

That said, there's still a lot of uncertainty as to what Nvidia has and hasn't done to Fermi. Efficiency in particular is an interesting word, as from what I've gathered it's far more efficent in terms of performance : peak FLOPS than GT200, but its shader count:FLOPS ratio is lower because they've removed a unit from the shaders that was underutilised in GT200. It'll be interesting to see what they've come up with.

 

[drunkenmaster]

The G80 was a huge step because it was the first unified shader architecture, don't expect that sorta jump with Fermi.
It HAS to be faster than a GTX295, otherwise it's in the same ballpark as a bandwidth limited 5870.

I really don't know why it wouldn't be, the very first review I checked, the 5870 review on Anandtech has the 295gtx consistantly 30% behind 285gtx's in sli when not cpu limited. Theres no reason at all it shouldn't be faster than a 295gtx, but not 30% ahead of it. People seem to think just over doubling the shaders will mean 100% performance improvement.

All of this is fairly silly, the thing is for right this second very few games push last gen performance hard enough to really warrant new £300 cards, TODAY. However the bigger/tougher games on gpu's this year should start to push them and the bleeding edge titles with dx11 might push them very hard indeed, its only with those titles, what both cards are really designed for, we'll know how they compare and how useful they are.

Though I'm in the camp that the sooner you get one the better because you end up with higher minimums(not just average) meaning smoother gameplay even if its not a huge difference in that many games. My point is just that in 6 months a re-review of the Fermi and 5870 could have a very very different conclusion.

I also think the clock speed issue and heat issue that hit Nvidia also played its part in the 5870, I imagine on a better process it would have launched at higher clock speeds. We'll really have to see exactly what speeds Fermi hits, shaders aren't everything if there are 512 but they are clocked 20% lower than their targets, well, we'll have to see what impact that has.

 

[asoxedri]

my bet is 5-15% faster then 5870 (g spot being around 10%), don't forget the royalty figures you'll be paying because its got a green badge on it! mind you same thing can be said about ATI lately lol

Tsk tsk, scaremongering ATI buyers

I'm safe anyways, can't go wrong with last gen graphics cards

what sort of temps you get under load and idle with your gfx? as soon as i purchased the asus 1, the vapor x came out. i was so gutted lol

 

[ETNiES]

Staying up to see some specs is a bit much lads.

 

[Hotwired]

what sort of temps you get under load and idle with your gfx? as soon as i purchased the asus 1, the vapor x came out. i was so gutted lol

56 degrees idle (fan @ 37%)
About 75 under load (fan goes to about 50%)

Ambient is about 20 degrees atm. Case cooling is on quiet.

 

[pinkacidbootson]

Staying up to see some specs is a bit much lads.

Im hosting a GTX 300 series spec reveal party, the guests should be arriving any minute now.

 

[drunkenmaster]

First info, supposedly from CES and being told to Charlie(yes i know, I'm not saying its certain) suggests that 512shader parts at slower than anticipated speeds are 60% ahead in the very best games(read 3 TWIMTBP games that are ultra Nvidia specific, not all of them) and 20-30% faster in others, I saw a think suggesting this before where it was 1.6x faster in Crysis and Hawx and only 1.2x faster in Fallout 3 and, I've forgotten the other one.

its about a 60% larger core, the 512shader part at higher clocks runs at 280W, for the SINGLE CORE version, opps. THE AIB's have been told to expect a handful if any 512 shader parts, the only parts Nvidia can ship in any numbers are going to be 448SP 1200Mhz parts(they had targeted 1600mhz). If the 512shader parts had the full 1600Mhz clocks, the 448 at 1200Mhz will barely beat a 5870, if the 512 shader parts were 1400Mhz, it won't be much faster than the 5870, barely at all in the worst games and 30% maybe in the best games.

The rumour is a single high end Fermi will cost around a 5970....we shall see.

The thing is, I can believe that, even from Charlie as those are roughly what we've been hearing everywhere for a while.

The question is where will it actually price a 512shader not availble part and the 448 shader lower clocked available part, at a 60% larger core(or more) its not going to be very close I'd imagine.

Tesselation sounds VERY interesting if he's accurate, not real tesselator hardware, but a sub type shader that can be used so theres a lot of raw power there. However Uniengine does really only tesselation so Nvidia apparently do VERY well on it, but in a real game where theres normal work to be done on shaders and Tesselation on top of it, you could see a massive hit on Nvidia hardware, that I'll be very interested in seeing how it works out.

I think unfortunately having 3 generations of an identical tesselator wasted in AMD chips for 3 years meant AMD weren't willing to go all in with a bigger/upgraded Tesselator, at a die space cost, only to see it canned again. If Tesselation takes off I can't see any reason, especially with a move to 28nm instead of 32nm that seems likely, that they can't massively increase the power of the tesselator hardware.

Its unclear if 512shader parts will be priced around the 5970, or the 448shader part, I get the impression the later which would be terrible if the 5970 beats the former in performance in most games. I assume Charlies putting bits and pieces together and probably didn't have an invite to the NDA covered launch.

 

[ejr22]

snip

No link? Or is this yet more rumour based on Chinese whispers started by gossip and hearsay?

Suppose there's not long to wait for official info anywho. I'll wait for that.

 

[drunkenmaster]

No link? Or is this yet more rumour based on Chinese whispers started by gossip and hearsay?

Suppose there's not long to wait for official info anywho. I'll wait for that.

Charlie, semiaccurate.com , I assumed it wasn't worth posting as 98% of the forum just assume he's wrong(yet fail to link to anything he's been wrong on when asked to).

He was at CES and his information isn't far off what we've been hearing, nor far from making sense and we know he talks to a lot of people in the business.

For reference he called the bumpgate situation, I don't think he did the fake Fermi's, called the dates to within a week of every stage of Fermi's developement over the last year, the 260-285gtx going EOL and a whole crapload of other stuff.

Out of all the leaked stuff out there, his is the very much most likely to be true, that doesn't mean it is as I said, nor that people haven't lied to him. But I wouldn't be remotely surprised if a lot of what he's saying is true.

However tomorrow isn't really a full launch either, I'd expect the 512sp part to be the focus, with very little info put out on the final specs of the 448 and very little mention on availability.

Raw specs of the top end card is about it, we're another 2 months from seeing real in store prices and availability and maybe what specs the only really available parts are. This launch could say 512sp's, 2Ghz, and $400, but in 2 months we only see $600 448sp 1.2Ghz parts

i'd quite happily buy a Fermi if at any stage with any combination I can get more performance for the same price out of it, 2x 384shader parts that cost less than a 5970 and beat it or something along those lines, I just can't see it happening at all.

[H] are saying their "review/preview" will be up at 11CST, which afaik is 5am here, so should be up in 10 mins, so my insomnia finally works to my advantage. However by their taster of their take on Fermi, they are impressed. But again the problem is what they launch, and what version is available to us, which we won't possibly know (unless Nvidia really ballsed up and some AIB's are very honest and on the record) till the cards are actually available. Does some like some hotlot wafers are being made already and we should see some cards mid march. But the review says they definately don't have hardware, no AIB's have hardware yet, so its not going to be sooner than March.

 

[drunkenmaster]

Reviews are up, [h] , Anandtech, and for the record Charlies numbers were up first, and correct.



The biggest thing [H] talks about from a quick glance is the "Awesome" Tesselation performance, however 60% faster, from a lot more die space for Tesselation than AMD from a 60% larger core seems not as impressive as it at first seems. As Charlie highlighted, the tesselation unit is scaleable on Nvidia, but only down, a smaller core will have less tesselating power while every chip AMD sells inthe 5 series has the same tesselator, meaning midrange and lower shader parts don't take a performance hit, nor do they lose performance when the shaders are doing other things, which is something Nvidia COULD certainly do, its very possible it will lose a huge amount of Tesselating speed in real gaming situations.

The good news is, Nvidia seem to think Tesselation is huge which means it should hopefully be in a LOT of games this year, and personally I think the IQ improvement it can(but won't always, good programming and design work is more important than the hardware) produce is fantastic.

The other good thing being if Tesselation kicks off big time, AMD can drop a bigger tesselator unit in, Nvidia would need to change their entire SP design to increase its power.

EDIT:_ I'll also happily admit the Architecture is a far bigger change than I thought, in some places good, in others not so good. They've stuck more in the Shader clusters, which improves speed of certain things, but they've also got an out of order items issue, which I think is somewhere they could really end up bottlenecked when doing the extra work of things like tesselation in real world with a lot of other data going in and out at the same time. But Nvidia have gone MASSIVE with the core, while taking a little bit of the AMD style of slightly different shaders added in to do different things design approach, but not to reduce size at all.

EDIT:- The bad news, for this launch Nvidia didnt' utter a word on clock speeds, power, cooling, yields, or much of anything useful it would seem. Which I gather from several reviewers is a pretty bad sign, Charlies info seems to be mainly from AIB's who would have that info and whose numbers tie in exactly with official info from Nvidia on the things they did release.

This might be a kick ass card(remember most AMD fans did say it would be faster), but if they could sell it as cheap as an AMD equivilent, wouldn't they tell us now so we could wait for March? If the 512sp part was going to be widely available and offer price/performance parity, or better, wouldn't they be shouting that from the rooftops.

I'm going to guess Charlie isn't very far wrong, a 448sp part the only thing really available, and it will cost significantly more and will probably be clocked quite a bit lower than the top end can't find it part. Which will mean significantly less tesselation power compared to a higher clocked 512sp part.


The really bad news, this "launch" tells no one a single thing about if Nvidia will offer a better buy than AMD come mid march, not even a tiny clue. A single card will only support 2 displays, the Fermi will need SLI to support 3 screens for surround gaming, which well, if they are cheap, won't be an issue, if they cost a bomb, makes it a useless feature to most of us

 

[muon]

No performance numbers though.

edit:

Okay, [H] have one gaming performance graph