3 40nm chips coming from NVIDIA

[Wreckage]

http://www.bit-tech.net/news/h...hips-coming-this-yea/1

We are ramping 40nm probably harder than anybody and so we have three products in line now in 40nm and more coming shortly," said Huang during the question and answer session following the earnings call. "We have the vast majority of their line cranking right now with new products. We are monitoring yields and they are improving nicely week-to-week-to-week and so at this point there's not really much to report," he added.

 

[imported_Shaq]

Very good news. I can't wait to see specs/benchmarks soon.

 

[thilanliyan]

I think some of this is also referring to mobile products.

From the original link:
"We say this because the general consensus is that one of Nvidia's first 40nm products will be a notebook part ? the GeForce 9400M, which is the same chip that is used in the Ion platform, has had some pretty good market penetration thus far."

Originally posted by: Shaq
Very good news. I can't wait to see specs/benchmarks soon.

My guess is you'll be waiting a while.

 

[SlowSpyder]

I am looking forward to the next gen chips from both companies for sure.

But I do find this quote a bit funny, "We are ramping 40nm probably harder than anybody..." Seeing as AMD has one 40nm mainstream part for sale vs. zero for Nvidia, it would seem that they didn't work quite as hard at getting 40nm out the door has he says.

I don't know enough about the manufacturing side of things, but Nvidia already has a very solid mainstream part with the 8800/9800/GTS250 cards, is it really that hard just to shrink them to 40nm? Or are they redesigning everything?

 

[Idontcare]

Originally posted by: SlowSpyder
But I do find this quote a bit funny, "We are ramping 40nm probably harder than anybody..." Seeing as AMD has one 40nm mainstream part for sale vs. zero for Nvidia, it would seem that they didn't work quite as hard at getting 40nm out the door has he says.

Its one of those quotes that of course is no-doubt technically correct once the caveats are rolled out in the fine print. "ramping harder"...maybe by NV's use of the term (and tense of it) ATI isn't "ramping" anymore, it already has ramped, past tense, and now they are in HVM phase. So between the two, NV is now the one that is ramping and thus, is ramping harder than anybody.

These kinds of twists on the words, spin-doctoring is the appropriate term, always have some manner of legal truth to them but of course they are worded so as to generate a false impression of something else being true to the reader.

Another technically accurate interpretation would be that NV believes they have more wafers (higher volume) of 40nm products swirling around inside the fabs than ATI does at the moment, so if both are ramping 40nm then maybe NV has twice as many 40nm wafer starts right now as ATI does, even though ATI is shipping the relatively smaller volume of wafers while NV is stockpiling enough good chips to hard launch or something along those lines.

It gets really hilarious when you are an engineer on the other side of those PR statements...I've seen my share of my employer's PR's where I read them and they directly relate to my project area and the fact versus fiction is just :shocked:

Bottom line is you have to just take these PR statements for what they are worth and consider who they are targeted towards. (shareholders/analysts/wallstreet)

Originally posted by: SlowSpyder
I don't know enough about the manufacturing side of things, but Nvidia already has a very solid mainstream part with the 8800/9800/GTS250 cards, is it really that hard just to shrink them to 40nm? Or are they redesigning everything?

It is that hard to shrink them to 40nm. Even if they merely wanted to keep the exact same architecture, xtor count, etc and simply shrink the chip to take advantage of the smaller dimensionality they could spend upwards of a year reworking the chip so it works properly and as desired/expected.

There's a reason we can count on one hand the number of companies doing this stuff at leading edge nodes.

 

[ronnn]

The 4770 seems to be sold out around here, which does make one wonder about yields.

 

[nitromullet]

Originally posted by: ronnn
The 4770 seems to be sold out around here, which does make one wonder about yields.

I was thinking the same thing. Maybe that's why NV is ramping while ATI is selling (or selling out of) 40nm chips. I'm sure that either tactic is a calculated risk.

 

[Keysplayr]

Originally posted by: nitromullet

Originally posted by: ronnn
The 4770 seems to be sold out around here, which does make one wonder about yields.

I was thinking the same thing. Maybe that's why NV is ramping while ATI is selling (or selling out of) 40nm chips. I'm sure that either tactic is a calculated risk.

I'm pretty sure Nvidia keeps and eye on things like this. Seeing low yields on even the smallest die (4770 is 137mm2? or 173mm2?). Either way, it's small. Which does not really bode well for TSMC's 40 (or it is really 45nm) process at this point in time and Nvidia could be waiting until it's yields on it's own 40nm products are sufficient to satisfactorily fill the channel. Or slowly stockpiling for release.

We don't have any 4770 40nm yield numbers, and I doubt we ever will, but if they are out of stock, it could be for 1 of 2 reasons. They are being snatched up like no tomorrow, or yields are low and stock is non-existent. From what we've heard about TSMC's probs, I vote the latter.

 

[Idontcare]

Originally posted by: Keysplayr

Originally posted by: nitromullet

Originally posted by: ronnn
The 4770 seems to be sold out around here, which does make one wonder about yields.

I was thinking the same thing. Maybe that's why NV is ramping while ATI is selling (or selling out of) 40nm chips. I'm sure that either tactic is a calculated risk.

I'm pretty sure Nvidia keeps and eye on things like this. Seeing low yields on even the smallest die (4770 is 137mm2? or 173mm2?). Either way, it's small. Which does not really bode well for TSMC's 40 (or it is really 45nm) process at this point in time and Nvidia could be waiting until it's yields on it's own 40nm products are sufficient to satisfactorily fill the channel. Or slowly stockpiling for release.

We don't have any 4770 40nm yield numbers, and I doubt we ever will, but if they are out of stock, it could be for 1 of 2 reasons. They are being snatched up like no tomorrow, or yields are low and stock is non-existent. From what we've heard about TSMC's probs, I vote the latter.

It's 137mm^2. Pretty good shrink, xtor density-wise. But yields...

 

[SlowSpyder]

Originally posted by: Keysplayr

Originally posted by: nitromullet

Originally posted by: ronnn
The 4770 seems to be sold out around here, which does make one wonder about yields.

I was thinking the same thing. Maybe that's why NV is ramping while ATI is selling (or selling out of) 40nm chips. I'm sure that either tactic is a calculated risk.

I'm pretty sure Nvidia keeps and eye on things like this. Seeing low yields on even the smallest die (4770 is 137mm2? or 173mm2?). Either way, it's small. Which does not really bode well for TSMC's 40 (or it is really 45nm) process at this point in time and Nvidia could be waiting until it's yields on it's own 40nm products are sufficient to satisfactorily fill the channel. Or slowly stockpiling for release.

We don't have any 4770 40nm yield numbers, and I doubt we ever will, but if they are out of stock, it could be for 1 of 2 reasons. They are being snatched up like no tomorrow, or yields are low and stock is non-existent. From what we've heard about TSMC's probs, I vote the latter.

TheInq (take it for what it's worth) had said, if I remember correctly, that yields for 40nm at this point are around 20%. I remember reading that on their site, I assume they were talking about the 4770. So it's going by my memory and from TheInq, so take that number for whatever you feel it's worth.

 

[Idontcare]

Originally posted by: SlowSpyder
TheInq (take it for what it's worth) had said, if I remember correctly, that yields for 40nm at this point are around 20%. I remember reading that on their site, I assume they were talking about the 4770. So it's going by my memory and from TheInq, so take that number for whatever you feel it's worth.

Yields change rapidly at this stage of a new node's introduction. 20% for NUBS (net units built) would be about right some 4 months ago.

Taken separately, functional and parametric yield are much higher, of course, but the product of the two is what is called probe yield and they aren't fantastic at this point in time once parametric yields (includes leakage binning) are factored in.

The challenge for NV, if they are producing GT300 on 40nm and if it is another 600mm^2 monster, is that the functional yield will be considerably lower at this point in time of the process maturity phase of TSMC's node.

40% functional yield on a 140mm^2 chip can translate into 5% functional yield on a 600mm^2 chip.

(this is part of the reason why the big-iron CPU's like Itanium, Niagara, and Power6 all lag behind the leading edge release timeline for a new node...until the node reaches entitlement D0 and functional yield there is little point in throwing wafers with 600mm^2 die at the fab)

 

[MarcVenice]

IDC, can you link me to 'where-ever' explaining functional/parametric/probe yield? Or write me an essay

 

[Idontcare]

Originally posted by: MarcVenice
IDC, can you link me to 'where-ever' explaining functional/parametric/probe yield? Or write me an essay

I learned it all on the job, but the simple of it is that functional yield are the chips that simply turn of/off (i.e. function) as they were designed to do per the electrical wires and transistors in them. (No open wires, no shorted wires, etc...defects are the predominate issue here)

Functional yield says nothing about how well the chip performs...it could run at 1MHz (and no higher) and be counted as functional, but you couldn't sell it. Or it could consume 1000W to operate at 500MHz, a sellable clockspeed but unsellable TDP.

That is where parametric yield comes in, the chip functions but does it function within a set of parameters that the market will accept? I.e. this is where "binning" comes into play.

Probe yield is simply the product (as in multiplication) of these two.

Probe = Functional x Parametric

Further yield loss can occur during dicing, packaging, and burn-in (not always done). At the end of it all you have what is called NUB (net units built) and NUB yield.

Functional yield can be improved upon by means of harvesting and having redundancy built into the chip's design, all of which comes at the expense of larger die and thus larger cost/chip.

Checkout this google book preview on yield

This pdf is also kinda lengthy but you can extract useful information from it relating to parametric yield.

 

[nyker96]

with all this talk of 40nm from nv, which models will be 40nm? will any of their current dx10 parts get 40nm makeover? or this is exclusively a dx11 thing for nv? and when will they intro these 40nm parts? June? July? aug? end of year?

after reading these comments from nv, I have more questions than answers.

 

[Genx87]

Originally posted by: nyker96
with all this talk of 40nm from nv, which models will be 40nm? will any of their current dx10 parts get 40nm makeover? or this is exclusively a dx11 thing for nv? and when will they intro these 40nm parts? June? July? aug? end of year?

after reading these comments from nv, I have more questions than answers.

I'd expect some lower end\lower margin products first to see the process improvement.

 

[Keysplayr]

Originally posted by: nyker96
with all this talk of 40nm from nv, which models will be 40nm? will any of their current dx10 parts get 40nm makeover? or this is exclusively a dx11 thing for nv? and when will they intro these 40nm parts? June? July? aug? end of year?

after reading these comments from nv, I have more questions than answers.

Hmm. If they do, it will be the first time (that I know of) that a GPU transcends more than two process nodes. If they shrink the G92 to 40nm, I think that will be a first. G92 would then have ridden along 65nm, 55nm and 40nm. Who knows. And of course, I could be wrong about the 3 node thing. I just do not recall any arch going longer than 2 nodes.

 

[Wreckage]

My guess is a mobile/intergrated chip, a GTX2xx shrink and something new.

 

[nyker96]

Originally posted by: Genx87

Originally posted by: nyker96
with all this talk of 40nm from nv, which models will be 40nm? will any of their current dx10 parts get 40nm makeover? or this is exclusively a dx11 thing for nv? and when will they intro these 40nm parts? June? July? aug? end of year?

after reading these comments from nv, I have more questions than answers.

I'd expect some lower end\lower margin products first to see the process improvement.

the fact that they acting all mysterious about their progress on 40nm parts aren't a good sign. They feel like amd right before the introduction of Phenom I and we all know how that debut turned out. if you ask me, they keeping so tight lipped because they are having some problems in this 40nm transition and doesn't want people catch the wind of it.