The G92 Architecture Rumours & Speculation Thread

Status
Not open for further replies.
So if 8800GT have 112SP , it will have 24 Texture units and 20 Rops ?

Who can explaine me whats the cluster number thing...

WRT TMU: basically yes. If Nvidia didn't change their TCPs in the meantime, that is. ROPs are quite independent from this as they are individually scalable.

TCP is btw the same thing as the "cluster number". It is aunit comprised of a 16xSIMD-ALU-block and a Quad-Texture-Memory Unit.
 
Does RV670 consider as mid-range or high end ?
All I can tell is it seems G92 performance depends on RV670 performance.

For manufacturers it is the asking price that defines the market segment.

Ergo, RV670 is mid-range, but offers higher mid-range performance. AMD can let it happen because their 2900 XT-s aren't exactly flying off the shelves, and they need to beat nVidia in as many market segments as possible since the high-end is owned by nVidia.

nVidia on the other hand has 3 high-end SKU's that sell very well, and could do so for some time to come. nVidia doesn't do anything for purely progress - progress takes a back seat compared to profit. So you can bet that they would not upset the lineup that makes them good money simply because technology-wise they are able to offer better performance for a lower price now.
 
AMD doesn't seem to mind that RV670 threatens to be faster than R600 and get up to GTX levels depending on the scenario. It has 76GB/s of bandwidth and faster clocks. That's a hell of a lot of card for $250.

So if Nvidia needs to compete with that I don't think they have a choice but to push for GTS-GTX performance in that price range.
 
For manufacturers it is the asking price that defines the market segment.

Ergo, RV670 is mid-range, but offers higher mid-range performance. AMD can let it happen because their 2900 XT-s aren't exactly flying off the shelves, and they need to beat nVidia in as many market segments as possible since the high-end is owned by nVidia.

nVidia on the other hand has 3 high-end SKU's that sell very well, and could do so for some time to come. nVidia doesn't do anything for purely progress - progress takes a back seat compared to profit. So you can bet that they would not upset the lineup that makes them good money simply because technology-wise they are able to offer better performance for a lower price now.


I would bet that G92 makes them more money (even at the lower price) then 8800GTS 320 ever did.
 
AMD doesn't seem to mind that RV670 threatens to be faster than R600 and get up to GTX levels depending on the scenario. It has 76GB/s of bandwidth and faster clocks. That's a hell of a lot of card for $250.

So if Nvidia needs to compete with that I don't think they have a choice but to push for GTS-GTX performance in that price range.

I'm sorry, but take a look at the RV670 tests so far (I refer to the HIS model). Under 4xAA and 8xAF it has trouble keeping up with 8800 GTS 320MB until 2560x1600.

http://www.hexus.net/content/item.php?item=10002

Basically, it is what it is - a chip with somewhat worse performance than 2900 XT (which rivals the 8800 GTS family itself). Definitely a good value for money proposition at higher mid range prices, but not something that would best even the 8800 GTS 640MB under typical usage scenarios (I think that playing at 1600x1200 or 1920x1200 without any AA or AF enabled is rather atypical these days ;) ).

Anyways, RV670 is not the reason why nVidia would have to sacrifice their geese that are laying golden eggs.

PS - where did you see that 2900 PRO has higher clocks than 2900 XT? The HIS model has 600MHz GPU and 1584MHz mem, definitely less than 2900 XT at 743/1650.
 
Who can explaine me whats the cluster number thing...


Cluster @ G80: 4 Double-TMUs (4TA+8TFs, ~free tri or 2x bi-AF) + 2 Vec8-ALUs (each 8 SPs)

Cluster @ G84/86: 8 TMUs (8TA+8TFs) + 2 Vec8-ALUs (each 8 SPs)

Cluster @ G9x: ???

ROP-partition: 4 ROPs + 64Bit memory-channel
 
erick, RV670 is *not* the HD 2900 Pro, and has not been released yet. RV670 will be released in November... You can't deduce its performance based on R600 benchmarks...

As for NVIDIA not doing things for "progress' sake" - yes and no. They won't reduce prices just for shit and giggles. But that won't affect their product roadmap at all: if competition is weaker, they'll price the same products higher, and that's all.

If NVIDIA or AMD can replace a chip by a new one on a lower process node to save money, they'll do it. If both do it at the same time, then logically they will both lower their prices accordingly; however, if only one does it, then prices will remain higher. In a highly competitive market such as GPUs, however, the former is more often true than the latter IMO...

As such, I still believe that G92 will (eventually?) replace G80 completely. I don't know if it will at launch, or even this year; but I don't believe in the G90 rumours, and there is no reason for NVIDIA to keep a gigantic 480mm2 die longer than they need to. Whether the GTX/Ultra will be replaced by a higher-end G92 SKU or the G92-based GX2 is the key question at this point really IMO. And as I have hinted at earlier, this depends solely on whether NVIDIA is willing to use GDDR4 this round.
 
Where did you see that 2900 PRO has higher clocks than 2900 XT? The HIS model has 600MHz GPU and 1584MHz mem, definitely less than 2900 XT at 743/1650.

Like Arun said, you're confusing RV670 with the clearance underclocked R600 parts (2900 PRO/GT). RV670 is rumoured to be hitting 825Mhz on the core with 2.4Ghz GDDR4 on a 256-bit bus. There have been no benchmarks as yet but since R600 doesn't seem to be benefitting much from its excessive bandwidth I expect RV670 has a very good chance of overtaking the 2900XT.
 
but since R600 doesn't seem to be benefitting much from its excessive bandwidth I expect RV670 has a very good chance of overtaking the 2900XT.

Doesn't seem to benefit from it compared to the G80 perhaps. But what happens when you downclock the RAM to get ~76 Gb/s on the R600 ?
 
I'm not quite sure if I'm reading AnarchX's link right, but the X axis is memory speed in Mhz right? If so, WTF? That's bizzarre...
 
Am I reading those graphs correctly?
Besides that bizarre dip, does the performance recover to near peak levels at about 500 MHz?

If that's the case, I'd expect that performance at a fully-clocked half-width bus to be better.
Downclocking is a double-whammy of reduced bandwidth and longer latency.
RV670 should hopefully avoid the latency penalty.
 
Very cool article. I wish they had an english version, but the graphs look pretty straightforward (assuming the x-axis is memory clock in all cases).

Boy are ATI getting a crappy ROI with their 512-bit bus. Halve the bandwidth and we're talking about a measly 5-10% performance penalty. Of course, it could be that going 512-bit doesn't really add much to the overall video card cost. Sixteen 8Mx32 chips vs. eight 16Mx32 chips isn't too much, is it? This isn't the console world where you worry about being pad limited in the future.

Looks like G80 could have been even faster if they went for 512-bit. With AA enabled it seems to be BW limited half of the frame time (5% FPS increase for 10% mem clk increase) in Q4 w/ AA at least. 512-bit and 512MB probably would have been cheaper than 768MB and 384-bit, too.
 
Last edited by a moderator:
I'm not quite sure if I'm reading AnarchX's link right, but the X axis is memory speed in Mhz right? If so, WTF? That's bizzarre...
The weird dip could be due to strange latency settings in the bios. I suppose some asynchronicity issues could pop up at a mem/core ratio that the chip wasn't designed for. It's only ~10%, after all.
 
A GPU company which acts "purely for progress" is a GPU company with a foot in the grave. These are for-profit corporations, not the United Way, or a national lab.
 
Looks like G80 could have been even faster if they went for 512-bit. With AA enabled it seems to be BW limited half of the frame time (5% FPS increase for 10% mem clk increase) in Q4 w/ AA at least. 512-bit and 512MB probably would have been cheaper than 768MB and 384-bit, too.

Yep, the GTX looks like it would benefit a lot from more bandwidth in that Q4 4xAA bench. Thanks for the link AnarchX!

But what's really going on with R600 though? The 4xAA graph in the Q4 comparison is basically a straight downward shift from the noAA graph. The noAA graph flattens out at 700Mhz implying something else is a big bottleneck at that point - texturing? Now if that's the case why the hell does the 4xAA graph flatten out at the exact same 700Mhz (at a lower performance level) if the workload should shift more onto the memory bus / shader core with AA enabled? :???:
 
Very cool article. I wish they had an english version, but the graphs look pretty straightforward (assuming the x-axis is memory clock in all cases).

Boy are ATI getting a crappy ROI with their 512-bit bus. Halve the bandwidth and we're talking about a measly 5-10% performance penalty. Of course, it could be that going 512-bit doesn't really add much to the overall video card cost. Sixteen 8Mx32 chips vs. eight 16Mx32 chips isn't too much, is it? This isn't the console world where you worry about being pad limited in the future.

Looks like G80 could have been even faster if they went for 512-bit. With AA enabled it seems to be BW limited half of the frame time (5% FPS increase for 10% mem clk increase) in Q4 w/ AA at least. 512-bit and 512MB probably would have been cheaper than 768MB and 384-bit, too.

If you only get 5% more FPS for a 10% Mem clock increase, then the card is not limited by bandwith.
 
If you only get 5% more FPS for a 10% Mem clock increase, then the card is not limited by bandwith.
no, the card is at least partially limited by memory bandwidth. do you think every frame (or every part of a frame) has the same bottlenecks?
 
More specs:
g92specqp7.jpg


G92 made @ UMC?
http://www.digitimes.com/bits_chips/a20071008PD218.html
 
Status
Not open for further replies.
Back
Top