[Shtal]

Hardspell released is list of possible specifications for the GeForce GTX 350 graphics processor (GPU):

* NVIDIA GeForce GTX 350
* GT300 core
* 55nm technology
* 576 sq.mm die area
* 512bit GDDR5 memory controller
* GDDR5 2GB memory, doubled GTX280
* 480 stream processors
* Grating operation units are 64 the same with GTX280
* 216 GB/s memory bandwidth
* Default clock speeds of core: 830MHz, shader: 2075 MHz, memory: 3360MHz (effective)
* Pixel fill-rate 36.3G pixels/s
* Texture fill-rate 84.4Gpixels/s
* DirectX 10, no DX 10.1 support yet.
http://futuremark.yougamers.com/foru...ad.php?t=85567

[AlphaWolf]

That'd be quite the trick if they could pull it off. Double the units and 50% faster shader clock in the same area with an optical shrink. I really don't know why they'd bother with 2GB of ram.

[AlStrong]

Is it just me or are the fillrates a little weird with the 830MHz core clock

[willardjuice]

576 sq.mm die area

They've learned nothing (if true).

[Anarchist4000]

Without simply killing off an entire chip I doubt we'll see the transition to a smaller chip happen right away. It's likely still to soon to have decided that massive monolithic chips were a bad idea.

[Florin]

Originally Posted by willardjuice

They've learned nothing (if true).

What should they have learned and when?

[CJ]

Originally Posted by Florin

What should they have learned and when?

Maybe that it's not always the size of the boat, but the motion of the ocean?

[dkanter]

Originally Posted by Florin

What should they have learned and when?

What should they have learned?

Don't produce >400mm2 chips unless you can sell them for >$1K.

Intel can get away with it because Xeon MP and IPF chips sell for $600 minimum, whereas an entire GT200 was $600 and included DRAM, heatsink, board, etc.

DK

[Florin]

Originally Posted by dkanter

What should they have learned?

Don't produce >400mm2 chips unless you can sell them for >$1K.

Intel can get away with it because Xeon MP and IPF chips sell for $600 minimum, whereas an entire GT200 was $600 and included DRAM, heatsink, board, etc.

DK

What is it with all the armchair experts here? Nvidia has been operating a highly profitable business for years and years. The low volume high-end segment halo effect is just one aspect of this.

Are you guys saying they should be looking to AMD for advice on how to run a business, just because they managed to pull off one promising videocard generation now?

[Panajev2001a]

Originally Posted by dkanter

What should they have learned?

Don't produce >400mm2 chips unless you can sell them for >$1K.

Intel can get away with it because Xeon MP and IPF chips sell for $600 minimum, whereas an entire GT200 was $600 and included DRAM, heatsink, board, etc.

DK

Yes, but did not people like PdM (was he really banned from RWT ?!?) and others report in the past those 400+ mm^2 IPF chips cost Intel like $150 (more or less) to actually make ?

(I remember an old RWT discussion about manufacturing costs of Itanium 2 chips)

[pjbliverpool]

Originally Posted by AlStrong

Is it just me or are the fillrates a little weird with the 830MHz core clock

To produce those fill rates (or close to them anyway) would require 44 ROPs and 102 texture units at those speeds.

I find that highly unlikely. I guess there could be other clock domains but I think its more likely that these specs are simply fake.

[Arun]

I really don't get the point of all this. The only lesson to be learned out of it is that GT200's architecture, relative to RV770's, sucks. And that it pays to have a better architecture and not underestimate the other guy.

The rest has very little to do with it, TBH (unless you think GT200 could be clocked at 800MHz if only intra-chip variability wasn't such a problem, which seems a tad extreme to me; and that's not the main problem anyway).

[CJ]

Originally Posted by Arun

....not underestimate the other guy....

And we have a winner. You can either choose to go for the washingmachine or for door number 3!

[Davros]

/Davros applauds No-X for puuting in the actual card names
eg: NV40 (GF6800)

[Humus]

Originally Posted by Shtal

* DirectX 10, no DX 10.1 support yet.

That'd be laughable if true.

[willardjuice]

Originally Posted by Humus

That'd be laughable if true.

What? The Nvidia focus group members tell me DX 10.1 is pointless.

[AlStrong]

Originally Posted by willardjuice

What? The Nvidia focus group members tell me DX 10.1 is pointless.

DX ten point one. nvidia am fail.

[Vincent]

Originally Posted by AlStrong

DX ten point one. nvidia am fail.

I heard that detail of Larrabee will be available in next IDF.

[no-X]

Originally Posted by willardjuice

What? The Nvidia focus group members tell me DX 10.1 is pointless.

Yeah, just like unification and HDR+MSAA

[trinibwoy]

Originally Posted by Humus

That'd be laughable if true.

Maybe but at this point I don't think DX10.1 is expected from them. If it's indeed the case that the base G8x architecture doesn't lend itself well to certain DX10.1 functionality there really isn't a compelling reason for them to make substantial changes until DX11.

[CarstenS]

Originally Posted by trinibwoy

If it's indeed the case that the base G8x architecture doesn't lend itself well to certain DX10.1 functionality there really isn't a compelling reason for them to make substantial changes until DX11.

I've always been curious as to what functions those really are? Nvidia - no wonder - is very tight lipped about it and no one ever mentioned which and foremost why DX10.1 should be giving them such a hard time.

[ShaidarHaran]

Chances this is real: zero

reason: die size is IDENTICAL to GT200

common sense...

[Shtal]

Originally Posted by ShaidarHaran

Chances this is real: zero

reason: die size is IDENTICAL to GT200

common sense...

GT200 @ 65nm 576 sq.mm die area
GT200 @ 55nm 470 sq.mm die area
GT300 @ 55nm 576 sq.mm die area

I was wondering about how would Nvidia be able to squeeze into 106 sq.mm die area - extra 240 stream processors for Total 480 stream processors into 576 sq.mm die area.


http://www.techpowerup.com/index.php?64193

[ShaidarHaran]

Always good to hear from you DK, thanks for stopping by! Can you comment any further on Paul's decision to leave? As stubborn as the guy was, I probably learned more from reading his articles on RWT than just about anywhere else (Johan de Gelas' work on Chip-Architect being the other).

Sorry for the OT, and thanks for the IPF analysis! Without Paul around to do it for us it's good to have someone else step up to the plate

Just a quick note on graphics card costs though:
You missed a couple steps of the chain, however - NV buys GT200 "kits" (DRAM, GPU, PCB, cooler) for perhaps $200, then sells them to AIB partners for perhaps $300, who then sell to resellers/distributors/OEMs for perhaps $400, who then resell for $450-$600. Note: all numbers are my rough estimates. I could be off by +/- $50 at any point of the chain.

[Rys]

Small note, but the die is 600mm^2, give or take, not 576.