NVIDIA GT200 Rumours & Speculation Thread

Status
Not open for further replies.
=>apoppin: I still don't think there will be a dual-chip GT200, even if the chip was shrinked. The architecture was modified to better cope with current and future games, but it still supports only DX10. So far, nVidia is doing a good job downplaying DX10.1, nevertheless, for BFUs 10.1 > 10.0. Now, imagine 45/40nm manufacturing won't be ready until Q1'09. By that time, Microsoft will probably have launched DX11 and it's not like nVidia to lag behind when a major DX version is available. Second, to allow for dual-chip cards, the chip would have to be shrunk to a size that would not allow 512bit memory. Not a problem in 2009, since there will be plenty of GDDR5 to go around, but oops, GT200 doesn't support GDDR5 or even GDDR4. That would mean having to desing a new memory controller, but I think nVidia will launch a completely new generation of GPUs instead. And make a lot of noise about DX11.

=>CJ: Since you already outed it, I can confirm that as well. (I also left a hint in the above text that somebody here could be able to see through, since I can't talk).
 
i like the idea of a dual chip especially since i bought a single pci-e 2.0 motherboard i have never liked the idea that you have to buy 2-3 cards to do a job 1 card should be able to handle especially since the cards cost 500-600$ im cheap i guess....
on another note 512 bit bus 1g memory that will be nice for my plasma 1080p crisis anyone?
 
I got the 240 streamprocessors confirmed today for GT200. As well as the 512-bit memory interface and 1GB memory.
So, do I get a cookie for breaking the news and being the first to point out the 240SPs possibility? :) (errr, why we're there... what about erasing from your collective memory how wrong I was about G92? :p)
Anyway every indicator I've got points in that direction too - as for RV770, on the other hand... Well, we'll see. Let's just say not all the rumours can be right, and I'm seriously starting to think something I've been suspecting for some time is correct... I'll refrain from mentioning it though, since: A) I'm not sure at all. B) I don't want to be as wrong as with G9x, thank you very much!
 
So, do I get a cookie for breaking the news and being the first to point out the 240SPs possibility? :) (errr, why we're there... what about erasing from your collective memory how wrong I was about G92? :p)
Anyway every indicator I've got points in that direction too - as for RV770, on the other hand... Well, we'll see. Let's just say not all the rumours can be right, and I'm seriously starting to think something I've been suspecting for some time is correct... I'll refrain from mentioning it though, since: A) I'm not sure at all. B) I don't want to be as wrong as with G9x, thank you very much!
Speak your mind, come on. We are always delighted when one of you guys open up your line of thinking, even if its a big tangent, the thought itself is pretty enlightening.
 
as for RV770, on the other hand... Well, we'll see. Let's just say not all the rumours can be right, and I'm seriously starting to think something I've been suspecting for some time is correct... I'll refrain from mentioning it though, since: A) I'm not sure at all. B) I don't want to be as wrong as with G9x, thank you very much!
Quit teasing like a lap-dancer and put out. :cool:
 
like call you names!!!!!!!!
or kidnap your fish (if you got one)
if not then another obscure pet
but seriously help your fellow man and share!
 
The 448-bit hint on the lower-rung GPU constrains things a bit.
I'm rushing the math, so anybody is free to correct or tell me it's already been confirmed:

If the assumption is that each cluster is tied to a fixed partition of the memory bus like before, the cluster count is either 8 or 16, since the memory bus width difference between 512 and 448 is 64.
Only 8 clusters with one disabled or 16 with two disabled works for this while also evenly dividing the 240 SP count.

The unfortunate side effect, though is that the SP count per cluster is either 15 or 30, which changes things...

If I've missed something or my math is screwed up, feel free to nitpick.
 
Know? Well, if we use the same definition, which is:
Dictionary.com said:
1. to perceive or understand as fact or truth; to apprehend clearly and with certainty: I know the situation fully.
Then I don't know anything... :p

As for doing 'not very nice things', I can make myself look dumb just fine, I don't need your help for that... ;) (and errr, no, this isn't schizophrenic, I'm just trying to imply I don't see the big deal)

Well, we'll see. - Duh.
Let's just say not all the rumours can be right - Duuuh?
and I'm seriously starting to think something I've been suspecting for some time is correct... - Uhh, yeah, and then it would turn out to be something ridiculous like 40 TMUs. Who cares?
A) I'm not sure at all. - No, really?
B) I don't want to be as wrong as with G9x, thank you very much! - Is that even theoretically possible?

Plus, I thought this was the GT200 thread - thanks for reminding us all, 3dilettante!
 
@3dilettante
There are two independent scalable parts at G8x/G9x and probably GT200:
- ROP-partitions: 64-Bit MC, 4 ROPs
- shader-cluster: 8 TMUs, 16 SPs /24 @ GT200

So GTX 280 seems to come with 8 RPs and 10 SCs -> 80TMUs, 240SPs, 32 ROPs, 512-Bit.

GTX 260 is speculated to be 7 RP and 8 SC. ;)
 
The 448-bit hint on the lower-rung GPU constrains things a bit.
I'm rushing the math, so anybody is free to correct or tell me it's already been confirmed:

If the assumption is that each cluster is tied to a fixed partition of the memory bus like before, the cluster count is either 8 or 16, since the memory bus width difference between 512 and 448 is 64.
Only 8 clusters with one disabled or 16 with two disabled works for this while also evenly dividing the 240 SP count.

The unfortunate side effect, though is that the SP count per cluster is either 15 or 30, which changes things...

If I've missed something or my math is screwed up, feel free to nitpick.

The ROP partition is tied to the memory bus, not the cluster.

In fact, G80 had 8 clusters and 6 ROP partition (64 bit wide)-> 384 bit.
G92 had the same 8 clusters and 4 ROP partitions -> 256 bit
G94 had 4 clusters and 4 ROP partitions -> 256 bit, too

So if GT200 has the same structure, it has 8 ROP partisions and 10 - 24 "SP" clusters (or 15 - 16 "SP" clusters), whereas a 448 bit part would have 7 ROP partitions and an undefined - indipendent - number of clusters (less than 10 or 16, anyway).
 
Well, no, in terms of perf/$ it actually turned out better than I thought. Anyhow, just so you can all laugh a bit at me, here's what I predicted privately on September 1st, 2007, with 24 SPs/cluster (i.e. 192 SPs/8C). That was, IIRC, assuming a die size of ~350mm² but I could be remebering wrong:

--------
G92 SKUs
--------
192-bit GDDR3 [8700 GS]: $189@384MiB/6C/700MHz [Late Q1]
192-bit GDDR4 [8700 GS]: $229@384MiB/6C/700MHz [Late Q1]
256-bit GDDR4 [8700 GTX]: $249@256MiB/7C/700MHz [LAUNCH]
256-bit GDDR4 [8700 GTX]: $299@512MiB/7C/700MHz [LAUNCH]
256-bit GDDR4 [8700 Ultra]: $399@512MiB/8C/825MHz [LAUNCH]
256-bit GDDR4 [8700 Ultra]: $499@1024MiB/8C/825MHz [LAUNCH]
256-bit GDDR4 [8700 GX2 Ultra]: $599@2x512MiB/8C/750MHz [Early Q1]
256-bit GDDR4 [8700 GX2 Ultra]: $749@2x1024MiB/8C/750MHz [Early Q1]

---

So yeah, I was basically expecting G9x to be the architecture refresh generation ala G7x, not GT2xx. Oops! And yes, back to GT200 now - I'll delete any post not about GT200 from now on. (in fact, I should have started doing that earlier, but heh)
 
thats not a horrible prediction based on their past i wish it came to fruition but cant get em all i guess :cry:

did somebody say how many cores will the gt200 have?
and i read about benchmarks all the time is it possible or are all of em attention grabbing fakes
gddr4.... why don't they use it availability what? its seems like a logical step for gt200 doesn't it?
but i guess they never got the prize for thinking logically:rolleyes:
 
The ROP partition is tied to the memory bus, not the cluster.

In fact, G80 had 8 clusters and 6 ROP partition (64 bit wide)-> 384 bit.
G92 had the same 8 clusters and 4 ROP partitions -> 256 bit
G94 had 4 clusters and 4 ROP partitions -> 256 bit, too

So if GT200 has the same structure, it has 8 ROP partisions and 10 - 24 "SP" clusters (or 15 - 16 "SP" clusters), whereas a 448 bit part would have 7 ROP partitions and an undefined - indipendent - number of clusters (less than 10 or 16, anyway).

Oh, Ok. I misremembered how those went together.
Never mind.
 
Status
Not open for further replies.
Back
Top