The Official G84/G86 Rumours & Speculation Thread

Considering the additional need for NVIO if they went with a G80-based part, I'm inclined to agree.

 

Meh. That cooler on 8800 is really very quiet and even lightweight. Way better than something with a little 5000 RPM fan on it.

 

Link

 

256bit !

 

Apparently someone made a mistake, because it makes absolutely no sense.

 

Seems like a card born to be vmodded.

 

Why is that? You think 64GB/s is too much bandwidth for the GTS? I tend to agree - since the 8800GTS makes do with 64GB/s it would be strange if they provided the same amount of bandwidth to the 8600GTS. Unless they are bumping up bandwidth for the entire line......

 

Or killing the 8800 GTS altogether...

Remember the 6800 GS vs 6800 GT situation ?
Both cards had the same bandwidth, and it was believed that the extra 4 pixel shaders and 1 vertex shader would give it the upper hand against the newer GS, but this last one had an extra 125MHz on the core, effectively cutting any real world performance advantage of the 6800 GT.

This "would" make sense if this 256bit, 512MB 8600 GTS was to replace the 8600 GTS in the 249~299 price bracket, accompanied by a sharp price drop of the 8800 GTX down to the former 8800 GTS 640MB level (~399 USD).
Of course, it's mere speculation.

 

This is one point: do not forget that G84 is only 25% of G80, so why should it need 76% of bandwidth?
The other points are, the small difference in transistorcount between G84 and G86 and price-segment in which the 8600-Series is set.

The 256Bit midrange-part I presume, we will see first with 65nm in end of 2007 or beginning 2008.
Maybe G94.

 

Yeah but the 8800GTS is a full 3x the configuration of the 8600GTS (ignoring the clock differential) - the 6800GS was a lot closer to the 6800GT and it was much easier to regain that performance by bumping up the clockspeed.

 

Between a 8800 GTS 320MB and a normal 8600 GTS 256MB there is a substantial difference, but the former is still far from being twice as fast.

A 80nm 8600 GTS with twice the bandwidth would certainly narrow that gap even further, while the lower production costs would eventually allow the card to come down to the 199~229 segment, which is out of reach for a 90nm G80 with a 320bit bus -and the extra NVIO chip-.


Most reviews I've read place the 8600 GTS as being a severely bandwidth-limited product, not core-limited.

 

Exactly.

8800GTS v 8600GTS
122% fillrate advantage, 100% bandwidth advantage

6800GT v 6800GS
Same fillrate, same bandwidth

Yup, wrong comparison.

I am more inclined to believe in a 8800GS (when the time comes; 64SP/192-bit) rather than a 256-bit G84.

 

I'm curious, why would you believe that it would be 64sp/192 bit vs something else with 256 bit? Because that would be effectively 1/2 an 8800 GTX in those areas? That is certainly plausible. However, note that a 192 bit bus would mean that they would have significantly different amount of on-board RAM vs a 256 bit variant: 192MB or 384MB RAM vs 256MB or 512MB RAM. I think it would be smarter for NVIDIA to go with the 256 bit bus variant in this price range, so that they don't get one upped in the frame buffer size and in bandwith.

One thing is for sure: NVIDIA has already set the expectation of 8800 GTS 320MB performance at the $250 price point. The problem is that the margins will not be so great with an 8800 GTS selling at this price point. So in order to have a successful transition, they will need to try to come up with a part that performs very similarly to the 8800 GTS 320MB, but one that is less costly to produce. Anything short of that will bring back bad memories of 9500 Pro getting axed in favor of the lower performing 9600 Pro. I think they could easily get there with a hypothetical "8700 GTS-type" 80nm card that has 64 sp's, higher clocked shader domain and memory than the 8800 GTS, 256 bit bus, and either 256MB or 512MB RAM.

 

Because the G80 being as large as it is, I expect there to be a significant amount of borked dies. Its been in production for almost six months now and we have seen only the 75% borked dies (8800GTS) and not the 50% ones yet. Think of an 6800XT in the Geforce8 series. I'm not sure what % of the dies will be not fully functional for one wafer of G80, maybe Arun would be a better person to even take a guess on that figure.

I think the 320GTS's market is $299-349 ($100 less) than the 640MB one. MSRP wise Nvidia doesnt have an SKU in the $249-299 range. Looking at the (rather) large performance delta between the 8800GTS & 8600GTS, they shouldnt have a problem slotting another borked GPU in there rather than designing one from the ground up.

 

And four ROP partitions?

 

A 256bit bus would imply that.
Anyway, remember that all of this could still end up to being a simple typo on the manufacturer marketing materials.

 

Isn't there a crossbar between the ROPs and memory controllers? Why are 4 ROP clusters necessary for a 256-bit bus? Or is the memory controller built into the ROP cluster?

 

ROP and MC are tied together 1:1. There's a crossbar with the ROPs/MCs on one side and the shader units on the other. ROPs are doing RMW operations (blending, depth/stencil test+update) where latency to memory is critical since fragments affecting the same pixel have to be serialized.

 

It's pretty close already:

http://www.newegg.com/Product/Product.aspx?Item=N82E16814133193

 

Gotcha, thanks!