Nvidia GT300 core: Speculation

xkcd is so awesome in so many ways

 

OMG it's alive.....

 

What happened... the mighty marchitects at the green team got scared and decided to go for a major redesign ?

That would leave AMD with a huge market window of opportunity to sell DX11 cards from October (= Windows 7 launch) until mid 2010.

See:

http://www.fudzilla.com/content/view/14104/1/

... quote mode on ..

Nvidia can't be too happy over the fact that AMD was showing off working silicon of its DirectX 11 hardware at Computex, especially as we heard a lot of rumours during the show that Nvidia won't have its own DirectX 11 products ready until sometime in the first half of next year.

We wouldn't even expect it to happen all that soon, as our understanding is that Nvidia is busy re-working a lot of its planned GPUs. From what we were told, TSMC is the source of this leak, as they've been talking to various people in the business and said that Nvidia won't have any DirectX 11 parts this year and not even early next year.

... quote mode off ..

 

If this rumor happens to be true, there's basically no conceivable way that this could be the reason. A reason that actually makes sense would be that there were some showstopping bugs in the silicon that require fixing.

 

...Or TSMCs 40G not being able to handle something like G300 in 2009...

 

Well, it might not necessarily be that, but even just an issue where the computer model of the processor layout didn't translate as well as hoped to the real process, such that a good design on paper failed to operate properly once produced.

Regardless, I think it's a good idea to bear in mind that this is just an unsubstantiated rumor at present, and thus not worth taking too much stock in.

 

G300 will need only two 6 pin power connectors. So the TDP must be under 225 watts.

http://www.hardware-infos.com/news.php?news=2997&sprache=1

http://www.brightsideofnews.com/news/2009/6/8/nvidia-gt300-targets-225w-tdp.aspx

 

So with a shrink to 32 nm, they may be able to do a quick x2 chip.

 

Why do they need a shrink? GT285 is presumably bigger and has a similiar TDP. On the other hand if they "need" an X2 then they're in big trouble....

 

Assuming the rumoured 495mm2@40nm are true for the "G300" you may combine that one with the fact that the TSMC 40nm yields aren't something to write home about at the moment. With a hypothetical less problematic process with far better yields and an by X% smaller die it's way easier to think of a dual-chip SKU; even more so that GDDR5 ram prices will have dropped further by that time meaning it'll be overall far easier to price such a monster better.

The picture gets even clearer if you think that "G300" might have 2GB onboard ram; a die shrinked dual chip SKU would have presumably how much ram even considering the possibility of one ROP partition being disabled?

 

Against "the largest Evergreen" in X2-config, they'll sure as hell (tm) [i.e. IMO] need a dual-gpu also.

 

Ouch... Tiles 'R US !

There might be truth in it, that Nvidia is not getting the yields and performance levels (incl. power) out of the 40nm node and are forced to jump to the half-node process (32 nm)

 

40nm is the half-node.

 

Not entirely true, since TSMC has not been offering 45nm for GPUs. So both TSMC 40nm and Globalfoundries 32nm can be considered full nodes.

 

As previously noted 32nm isn't a half node, while 28nm is afaik. In any case if they'd go for such a scenario they'd give AMD IMHLO a release advantage of at least 6 months and I don't consider NV to be as dumb.

 

The 40nm node is a refinement of the 45nm process with the same equipment and same general process.
I believe TSMC still considers 40nm as a half-node, even if it deemphasized 45nm.

Was 40nm better than 45nm? If so, in light of the current troubles, ick.

 

this whole 'half node' and 'full node' stuff is a bunch of marketing crap. 40nm is a small optical shrink of 45nm. It might be they have different strain techniques available, but if you look at things like contacted gate pitch and the manufacturing process it should be clear.

Also, I don't know why anyone would take Fudo quoting rumors seriously. ISTM that NV will probably have their DX11 part out in 1Q10.


DK

 

ISTM that both IHVs are struggling if possible to release at the same time as the win7 release (or damn close to it) and there's a shitload of flip flopping rumours from various websites that A will arrive before B and vice versa.

A couple of weeks ago we heard that AMD's X11 GPU is supposedly "delayed" and won't arrive in July and NV will arrive first while nowadays it's the exact opposite.

I refuse to read into any of that crap personally; I'm pretty sure that both IHVs when asked will claim that they're "on track". If there's a problem NV won't be able to hide it for very long.

 

Actually TSMC considers 40nm and 28nm to be full nodes, but as dkanter nicely points out that's just pure marketing. It does mean that there's an entire ecosystem around it for things like RF, whereas this obviously wasn't the case on 55nm, but that's about it IMO.

Remember there never was a 45G process, only 45LP (and Qualcomm was actually AFAIK the exclusive customer on that). And I have yet to hear anything indicating that 45LP and 40LP aren't both fine in terms of performance/power/etc. (although it turns out there is no performance benefit over 65LP, only power and density) - maybe not yields though, but like every generation that's mostly a question of time.

 

It seems NV is not able to hide it. CJ said the same thing in the R8xx thread the last week. Time will tell if GT300 is going to be R600 all over again.