Fermi A3 silicon is in the oven, Real launch possible in February!

[josiahsuarez]

http://www.semiaccurate.com/2009/12/10/fermi-a3-silicon-oven/
IT LOOKS LIKE Nvidia has put the A3 stepping of Fermi in the oven, it happened some time last week. That means that you may see cards as early as February, depending on two or three other factors.

If you assume that A3 went in the oven on Dec 1, that means Nvidia will likely get the first hot lot silicon back in house on Jan 1, 2010 give or take a few days. It looks like Nvidia testing staff are going to have very happy holidays. In their cubes. What does that mean? Sit back and enjoy the numbers.

If you assume that A3 will fix all the problems, and that it will be the launch stepping, there are two main things that will affect the launch date from there. First are the risk wafers. If they are still valid, then that will probably shave a few weeks off the time until production silicon will start rolling off the lines. Lets call the savings about four weeks, or six weeks versus ten weeks if the risk wafers had to be scrapped.

The next variable is based on how sure Nvidia is that A3 worked out the way it wanted that to. If it is 100% sure that A3 is the one it will launch, Nvidia could have started running wafers at the same time it put the A3 hot lots in.

Since the problems that we hear about on A2 silicon are related to bin splits and clock speed problems, and general yield in the good versus bad chip sense, this might be a fairly large risk for Nvidia to take. Given that, six weeks out from Dec 1 if it put A3 in on day one and the risk wafers are still valid, ten weeks if not. If it waits, that is +6 and +10 weeks from Jan 1 or so.

Last we have the time it takes to make boards, put them in boxes, and ship them. A good guesstimate is two weeks for the initial boards, including air freighting them from Taiwan or mainland China to get them ready for the big party. This assumes the prep work has been done, and the PCBs are simply waiting for chips.

The first take home message is that Nvidia can likely show hot lot early A3 silicon at CES, but we doubt the company will launch it there. Tegra 2 is probably a better fit for that, but don't be shocked if it has A3 silicon floating here and there.

If all goes well, risk wafers valid, A3 production went in on Dec 1, and all the stars align, you are looking at Dec 1 +6 weeks +2 weeks, or about Feb 1 for real availability. More realistically, if it waits until the A3 hot lots get back, that would be Jan 1 plus eight weeks, or March 1 for a hard launch. Basically, the best case is Feb 1 for a real launch.

The negative view is that the risk wafers are invalid, and Nvidia waits for A3 to come back before it puts in production wafers. That would be the outer bound, Jan 1 +10 +2 weeks, or April 1, Q2 of 2010. If you have a really negative outlook and think there will be an A4 stepping, add about four to six weeks for every stepping past A3. That is really unlikely to happen - really, really unlikely.

Toss into the ring that multiple Nvidia partners are telling SemiAccurate that they have been promised silicon in February, with volume some time after that. Realistically, we would estimate mid-February for real parts hitting the shelves, in VERY limited volumes.

The last open question is what A3 will be. We are told bin splits for A2 would put the top clock for Fermi at about 500MHz, and yields were described as "alarming". Last spring, insiders told us that initial targets were 750MHz , 50% up on A2. Nvidia PR drones were, note the past tense, claiming that it would beat Cypress by 40%, so a 500MHz version is likely to be a tad slower than an ATI Radeon HD5870. Not a convincing launch message there.

A3 will almost assuredly up the clocks quite a bit, but where they end up, and what percentage ends up in the top bin will be the make or break questions. We will know when A3 silicon comes back in a few weeks.S|A

[lee63]

Finally some good news...now lets see some #s

[Redeemed]

@ bolded... don't expect it to be cheap- or priced competitively, unless nvidia is will to sell them at or below cost.

[TheANIMAL]

@ bolded... don't expect it to be cheap- or priced competitively, unless nvidia is will to sell them at or below cost.

Prices must be competitie other wise people will only buy HD5xxx's.

[josiahsuarez]

assuming it's faster than 5870 and slower than 5970, it would have to be priced accordingly. of course a lot would depend on 5890 and it's price/performance/launch date.

[lee63]

I think NV need to be careful on this one...they must be competitive otherwise no one will buy the cards.

I really don't see the fastest single chip card being more than $400, $450...of course there will be price gouging if yields are low but I'm talking about MSRP.

thoughts ?

[Thunderbolt56]

If any single slot card is more than $450, it better add 2" to your peen, whiten your smile and make you irresistible to the opposite sex.

Smart people just won't pay more than that anymore.

[josiahsuarez]

if it were in fact 40% faster than 5870, it would make a certain amount of sense to be about 40% more expensive which would put it around $550. after all 5970 is about 50% faster and it's $600, so that seems to be the rule. but they might not make the 750MHZ target, so it's hard to assign a value to it without knowing what performance is like.

[a12ctic]

The fermi is more than three times as big as the r800 chips, not only that, but it uses a much more expensive memory interface. AMD can afford to sell the r800 for way cheaper than Nvidia could ever hope to sell the fermi. Unless this chip kicks some serious ass, Nvidia is going to be in for a bad year.

[Heinz68]

I think NV need to be careful on this one...they must be competitive otherwise no one will buy the cards.

I really don't see the fastest single chip card being more than $400, $450...of course there will be price gouging if yields are low but I'm talking about MSRP.

thoughts ?

It's hard to speculate on price without knowing the performance. If the Fermi (GTX 380) is 40% faster than HD 5870 as some of the NVIDIA PR were claiming in the past according to Charlie article, than the Fermi would compete with the ATI HD 5970 and mostly likely be prized as such, which means close to $600.

If the above is the case than the GTX 360 is going to be as fast as HD 5870 and compete with it in the same price range which means about $ 400.

The other possibility is that ATI will have faster HD 5890 ready at Fermi launch and reduce the prices on all cards which will also keep the NVIDIA Fermi cards priced lower.

I don't think NVIDIA can afford to start any price war with bigger more expensive chip and if the performance is there they might not need it.

PS
Times and things change, I'm surprised that this time nobody is upset with the Charlie article.

[Redeemed]

That's kind of my point.

nVidia cannot under sell AMD- they're chips cost less to produce, thus they can sell them lower while still making a profit.

nVidia doesn't have a choice but to sell them at a reasonable price... they cannot win a price war here due to how expensive their product is to produce. And with a very limited supply at first- simple laws of supply and demand come into play here- I'm rather positive they're going to be expensive at first.

[Ninja Prime]

if it were in fact 40% faster than 5870, it would make a certain amount of sense to be about 40% more expensive which would put it around $550. after all 5970 is about 50% faster and it's $600, so that seems to be the rule. but they might not make the 750MHZ target, so it's hard to assign a value to it without knowing what performance is like.

Don't plan on them making that target. Top end Tesla products with Fermi chips are clocked at 615mhz, with a TDP of 225W. They can't really go much higher, maybe 650mhz, without going into insane heat/power levels. I'm fairly sure that they won't launch much if any higher than 650.

[LydianKnight]

as they were targeting 225W that's probably the final target clock, somewhat around 615, even I'd say more like 600 and a bit below... maybe even 576MHz like in some GTX295 models (even my GTX260 goes 576MHz)... sounds 'reasonable' at this point.

[XMAN52373]

The fermi is more than three times as big as the r800 chips, not only that, but it uses a much more expensive memory interface. AMD can afford to sell the r800 for way cheaper than Nvidia could ever hope to sell the fermi. Unless this chip kicks some serious ass, Nvidia is going to be in for a bad year.

WHAT?! Where the hell do you get that fermi is 3x as big as R800? R870 is 334mm squared, Fermi is reportidly in the 490-550mm squared range. That is now where near 3 times the size.

[XMAN52373]

Don't plan on them making that target. Top end Tesla products with Fermi chips are clocked at 615mhz, with a TDP of 225W. They can't really go much higher, maybe 650mhz, without going into insane heat/power levels. I'm fairly sure that they won't launch much if any higher than 650.

It wouldn't surprise me if they did.

[Ninja Prime]

It wouldn't surprise me if they did.

...and why is that? You think they have some magic ninja wizards on their team? :p

Really, though if a Tesla board is hitting 225W at 615mhz, and the highest TDP of a single chip card ever launched was the GTX280 at 236W(and remember all the overheating problems they had, at that), then they can't really go much higher than 615. Maybe up to about 650, unless they have some magic way for GPU cards to take less power than Tesla cards, which could be possible for certain features to be disabled on the GPU versions. However, the few features you could turn off won't make up very much power, you're still looking at over 200W. You might get +10% on top of that, but barely.

[Ninja Prime]

Most likely the consumer end cards for 3D won't haeve the ECC enabled and maybe not have the 64bit FP enabled. Nvidia said they can turn off parts of the chip that don't benefit games. By turning these off the chip can run cooler and be clocked higher.

Yeah, thats what I'm refering to, I just don't think those options are all that much of the chip, probably less than 5%.

[TheANIMAL]

WHAT?! Where the hell do you get that fermi is 3x as big as R800? R870 is 334mm squared, Fermi is reportidly in the 490-550mm squared range. That is now where near 3 times the size.

334mm squared is 111556 mm2 area

550mm squared is 302500 mm2 area

302500/111556 = 2.7

So he was exagurating, but only by a little bit.

[XMAN52373]

334mm squared is 111556 mm2 area

550mm squared is 302500 mm2 area

302500/111556 = 2.7

So he was exagurating, but only by a little bit.

Nice to see basic math skills have gone to **** in this country.

550/334=1.6 times the size.
490/334=1.46 times the size.

[Woodelf]

I still think they went back to the drawing board, once they saw the 58xx launch.
There are alot of good point's here on the board. In definitly don't want to be in nvidia's shoes right now since every day they are loosing consumer faith. They're fan base should be feeling a little bit neglected (i am) from the lack of encouraging news, or even a high five. I mean they are about as anal of a company toward's those who by now, in nvidia's best interest (if they care what the high enders and share holders think at this point) deserve a very good reason not to switch over. Benchmarks should have been released to prove that they deserve our holding out to the new year. A2 silicon wasn't good enough, but couldn't they at least get us some benchmarks to give us hope on it. I fully intended to spend at least $600 on the nextgen by Dec 30th. I feel there has been justifiably enough time to persuade me not to go ATI again (if nvidia even cares, as I'm feeling right now that they don't). Basically the only thing I hear that sounds appealing is that the card is smaller and an expired promise of 30-40% faster than 5870 speeds. Although it does kind of sound cool to know YOU have two 12+ inch quad cards in your case too.
Dual 5970's are the best dx11 you can get. I have not heard that nvidia is going to 30-40% beat that. I wan't to hear that they benched better, have more memory and will be fair priced compaired to dual 5970's. Nothing else will do. I could give a rat's a$$ about cuda or any other scientific crapola. I'm a gamer.

Please don't ban me for this drunken rant.(lee)

[Ninja Prime]

334mm squared is 111556 mm2 area

550mm squared is 302500 mm2 area

302500/111556 = 2.7

So he was exagurating, but only by a little bit.

You're mis-reading the spec. 334mm squared means 334 square millimeters, its already an area measurement. So its only about 60% ish bigger. Still huge nonetheless, but not quite GTX280 huge.

Think about it for a second... 334mm is 33cm, do you really think the chip is over one foot by one foot? ;)

[josiahsuarez]

the chip is over one foot by one foot?

no wonder the yields were low! :O

[XMAN52373]

You're mis-reading the spec. 334mm squared means 334 square millimeters, its already an area measurement. So its only about 60% ish bigger. Still huge nonetheless, but not quite GTX280 huge.

Think about it for a second... 334mm is 33cm, do you really think the chip is over one foot by one foot? ;)

when talking about CPUs or GPUs, you refer to the die space package, IE its width XX.xmm x XX.xmm length, which is referred to in terms of mm^2(squared). SO for Cypress, its 334mm^2 and Fermi 490-550mm^2. I simply wrote out squared as I couldn't make the 2 appear to the upper right of the seconded m.

[Xion X2]

when talking about CPUs or GPUs, you refer to the die space package, IE its width XX.xmm x XX.xmm length, which is referred to in terms of mm^2(squared). SO for Cypress, its 334mm^2 and Fermi 490-550mm^2. I simply wrote out squared as I couldn't make the 2 appear to the upper right of the seconded m.

My god....

[XMAN52373]

My god....

Ina phrase, WTF is your issue? go back into your cave.