ELSA hints GT206 and GT212

Ok this is staring to get confusing,
first we has pipelines,
then we had pixel shaders and vertex shaders,
then we had sp's,
now we have tcp's.
what the bloody hell is a tcp ?

He meant thread (or was it texture?) processing clusters, TPC.

I blame the marketing departments. They should be put to work cleaning the Nvidia and AMD headquarters hallways instead.
 
Sorry for the accidental flip key-stroke, it's a TPC (Thread Processing Cluster) and there is actually nothing to do with the texturing, at first place. ;)
 
They need something less complex than a GTX280, but faster, with 8800GT-like qualities, e.x. lower power consumption and less noise (and this time, good yields, too).

Which is exactly what I think the roadmap shows. It says GT206 and GT212 and not GTX anymore.

Actually I don't think that's the case. It makes a lot more sense to think that GT212 and GT206 refer to the codenames for the chips and not the actual card names.

For example GT200 -> GT206 and GT212

Read of that what you will, but it seems like they will be designing seperate chips to replace GTX260 and GTX280 rather than using a single chip and disabling parts as was the case with GT200, G92, G80...

Regards,
SB
 
=>Silent_Buddha: They can still disable parts of the chips and I believe they will still do that with GT206 and GT212. Even though it doesn't seem so, it is cheaper than designing two chips with similar performance. With GT206, nVidia will have a GPU with better price/perf ratio, but still not a direct competition for the RV770. I pretty much think that's where cut-down versions come in. The GT212 will probably be another monster GPU where cut-downs are a necessity more than a marketing decision.
 
Actually I don't think that's the case. It makes a lot more sense to think that GT212 and GT206 refer to the codenames for the chips and not the actual card names.

For example GT200 -> GT206 and GT212
AnarchX already pointed that out and I agree with that of course.

What I basically was trying to get along is that NV´s next ASICs GT206 and GT212 need to be a lot more cost-sensitive, while both performance should be higher and power consumption should be lower.
 
The codenames suggest that it isnt a new architecture by any means.

Any ides why the refresh of GTX280 is at 40nm and GTX260 at 55nm?

And erm first post? :D
 
=>Cookie Monster: GT206 will be on the market sooner than GT212, so that's why. 40nm still feels far away for some reason. nVidia wants to gamble and bet on the new technology, but they still need some good produts in the meantime.
 
Hasn't TSMC announced that 45GS will start production in the second half of 2008? The answer why NV or any other IHV wouldn't present anything before early 2009 is pretty obvious.
 
With GT206, nVidia will have a GPU with better price/perf ratio, but still not a direct competition for the RV770.
Why not? If it's 8 TPCs @ 55nm without FP64 (CUDA 1.2) with less on-chip memory and with 256-bit GDDR5 bus i'd say it'll be exactly the direct RV770 competitor which could probably be used by NV for GX2-type AFR-board for some "direct" RV770X2 competition as soon as they'll have enough chips.
 
=>DegustatoR: GT206 is supposed to have 9 TPCs (216 SPs) and will probably run at higher clocks than the current GTX 260. RV770 can't compete with that.
However, GT206 could come close to RV770 in terms of manufacturing costs. If, as you propose, nVidia chooses a 256bit bus, removes the FP64 units and perhaps removes some redundant parts of the chip (that are in GT200 to improve yields), the result could have around 1B transistors, while being faster than RV770 and there would be the possibility of a GX2 card. Whether that's what nVidia wants, I do not know...
 
=>DegustatoR: GT206 is supposed to have 9 TPCs (216 SPs)
Doubtful. Why would they chop only one TPC off GT200? It's not that much of anything -- die size, perfomance, etc. Such a chip @ 55nm will compete with GT200 which is bad since GT200 will still be in production as the only chip with FP64 from NV.
Plus i feel that NV won't go for an odd number of TPCs at all. An 8 TPC chip will be more differentiated from GT200(b?), it'll be simplier and more inline with what AMD has to offer for the time being -- it'll probably have much higher clocks than GT200 in GTX260 and should end up being faster in general than RV770 in 4870 even with 8 TPCs.

However, GT206 could come close to RV770 in terms of manufacturing costs. If, as you propose, nVidia chooses a 256bit bus, removes the FP64 units and perhaps removes some redundant parts of the chip (that are in GT200 to improve yields), the result could have around 1B transistors, while being faster than RV770 and there would be the possibility of a GX2 card. Whether that's what nVidia wants, I do not know...
If that's what NV wants (to have a direct competitor to RV770 and a basis for a new GX2 card) then having 8 TPCs and higher clocks makes more sense than having 9 TPCs.
 
I'm not following your logic. First you say that 1 TPC is "not that much of anything -- die size, performance, etc." Then you say "having 8 TPCs and higher clocks makes more sense than having 9 TPCs."

If 1 TPC is "not that much of anything" then why couldn't you have 9 TPCs with higher clocks?
 
If 1 TPC is "not that much of anything" then why couldn't you have 9 TPCs with higher clocks?
Well you obviously can't have both.
To compete with RV770 NV needs to lower the die size, not to improve performance. So they should chop as much as they can off GT200 while maintaining ~GTX260 performance. From this POV going 8 TPCs with higher clocks is preferable to 9 TPCs with lower clocks.
 
An 8 TPC chip will be more differentiated from GT200(b?)
Well that's the catch, there won't be any GT200b. Besides, we don't know if nVidia removed the FP64 logic from GT206. Maybe they didn't, and GT206 will offer performance similar to GT200 (because of higher clocks on 55nm).
DegustatoR said:
So they should chop as much as they can off GT200 while maintaining ~GTX260 performance.
I think they don't want to do that since such a chip would be too close in performance to G92b. Then, once you take the defective chips and deactivate some units, you're at G92 performance and you don't want to have two of your cards competing each other. You want the cut-down GT206 to still have its own segment.
 
Well that's the catch, there won't be any GT200b.
Then it means that GT200 will be used in Quadro/Tesla only.

Besides, we don't know if nVidia removed the FP64 logic from GT206. Maybe they didn't, and GT206 will offer performance similar to GT200 (because of higher clocks on 55nm).
I think they did. FP64 is flawed even on GT200, it'll be even worse on any slower chip. The way i see it GT206 is price first, features and performance later.

I think they don't want to do that since such a chip would be too close in performance to G92b. Then, once you take the defective chips and deactivate some units, you're at G92 performance and you don't want to have two of your cards competing each other. You want the cut-down GT206 to still have its own segment.
I doubt that any 256-bit GDDR5-card will ever be close in perfomance to G92b -)
But if we're settling on no GT200b at all and GT206 as a GT200 replacement for the mainstream then i really don't see the point in choping 1 TPC off GT200 -- why not leave it as it is, 10 TPCs @ 55nm with 256-bit GDDR5 but with less redundancy, less on-chip memory and no FP64? 1 TPC isn't big enough to seriously improve anything comparing to GT200...
 
Don't ask me, I did not make these numbers up. But once they are redesigning the chip for narrower memory controller, GDDR5 support and removing the FP64 logic, they could very well chop off a TPC. Why is it necessary to stay at GT200 numbers, because of nostalgia?
 
Back
Top