G92?

J

Jadestar

Weaksauce
Joined
Oct 6, 2002
Messages
124
A few questions.. I know that we have very little information, but I think you guys are better at analyzing what information is available than I.

1) I've seen articles commenting on the G92 nearing 1 teraflop.. I really have no idea what that means, but people seem to think it's a good thing. The thing is, I've seen other articles mentioning that the G92 is built off of the G84, (like this one) and thus would fall short of exceeding the 8800GTX in performance. That doesn't really make sense to me, but I've seen that information linked multiple times. How's that going to work? It should seem obvious that the G92 will outperform all current G8X cards, why would the case be otherwise?

2) 65nm.. what does this mean? =)

3) Is there any chance that the G92 cards could be a little less power-hungry than the current cards? Or at least in the same ballpark? I got a 550w PSU for my new computer, and at this point I'm really stuck with it-- I'm kinda leaning towards using a temporary card and waiting for the G92 instead of getting an 8800GTX now-- will I be able to? A high-quality 80%+ 550w PSU should be able to power a single G92-based card, wouldn't you think? I mean, obviously none of us know, but you guys could make an educated guess that would be much more... well, educated, than any guess I can make.

Okay. I guess that's all. I'm still trying to feel out my options.. I think I'd like to wait for the G92, I just want to be reassured that I'll actually be able to use it, and that its performance will be worth the wait.
 
1) G92 might not be the high end
the GTX us G80 and 7800GTX is G70
2)die/fab size
3)almost certainly
smaller area
 
The 65nm part is the size of the components that make up the GPU. Smaller compenents mean less voltage required to push electrons through the circuit (due to shorter distances). Lower voltage mean less resistance. Less resistance means less heat. That is of course assuming nothing else is changed inside the core. The Prescott P4's are the anti-case to what I just said, but that has a lot to do with the radical changes made to the P4's architecture (twice the cash, 11 more execution stages, etc) and the poor thermal design of the s478 platform.

To get a image on what 65nm would look like: Imagine you have a 1 meter long stick. Then cut it up into a billion (literally) pieces. Grab 65 of them and put them back together. There you have a 65nm long peice of the stick. :)
 
a flop is a measurement of raw calculations persecond, but I'm not sure exactly what a flop is. 1 tera is 1,000,000,000,000 of the base units.

In the past two series, the GX0 has been the first high performance part, the GX1 is the 2nd generation highest performance part (and ultimatly the highest performer from the series), the GX2 is the mainstream performance part, and the GX3 is the lowest end part. ex, the G70 is the core on the 7800GTX and GT, the G71 is the core on the 7900GTX, 7950GX2, and 7900GT, and the G72 is the core on the budget (7200s, 7300s) components, and the G73 is the core on the mainstream (7600s) components.

a nanometer (nm) is 1/1,000,000,000 of a meter, and is how a TSMC, Intel, and AMD measure the distance of their components. the person above explained the advantiges.
 
It should be noted that flops is not a clear indication of gaming performance. For example, the 8800GTS has a peak performance of 345.6 GigaFlops (~345.6 Billion Floating Point Operations Per Second), while the 2900XT has peak performance of 475 GigaFlops. Clearly even an increase of 130 Gigaflops does not entirely show in the results. While I don't want to start yet another one of those threads, operations per seconds is not a clear indiction of gaming performance.
 
A flop is a FLoating point OPeration. A floating point number is any number from 0 to less than 1 (e.g 0.75 or 0.9939329353322). An operation is adding subtracting, multiplying, dividing, etc. So 1 teraflop would mean that the processor can perform 1 trillion operations every second. Now the benchmakr in and of itself means very little. You have to take into consideration how many OPs are need to complete whatever procedure is being done. That in turn depends on the efficiency of the processor.

For example. Processor X can perform a given procedure (set of operations) using 5 discrete operations to complete and it can only do 75 OPs/second. Processor Y has to perform 10 discrete operations to complete, but it can do 100 OPs/second.
So, Processor X can perform that requested procedure 15 times every second. Processor Y can perform the given procedure only 10 times for every second even though it does more individual operations per second.

Hopefully I didn't mangle that up too much. It's late and I'm tired. If I did, someone please straighten it out.
 
Thanks for the explanations. =)

So what I'm getting from this thread is that we have almost no indication whether the G92 will be high-end, low-end, etc. within nvidia's next generation of cards? I guess it wouldn't do me much good to wait for it if it's not going to clearly outperform the 8800 series.

Weird.
 
Jadestar said:
Thanks for the explanations. =)

So what I'm getting from this thread is that we have almost no indication whether the G92 will be high-end, low-end, etc. within nvidia's next generation of cards? I guess it wouldn't do me much good to wait for it if it's not going to clearly outperform the 8800 series.

Weird.
Click to expand...

Nobody has any indication of anything yet. Nvidia is very tight lipped about upcoming products. In fact I think the 8800 surprised every one with how well it did. Everything we see or read about G92 is just people trying to guess the minds of those who do know but can't legally talk about it. Nothing more.

As far as waiting, don't. If you need/want something now get it. Otherwise you'll always be waiting for the next best thing because as soon as G92 is release rumors of G10x will start flooding the net as the next best thing.
 
I don’t mean to derail this topic but will you require a PCIE 2.0 motherboard to run these cards?
 
Typically the x0 product is the high-end, and the x6 or x4 or x2 are the midrange and value.

Just guessing, G92 sounds like midrange.
 
As it is the case with all rumors and speculation at the moment, I'll still wait for either more information on G92 to come out or till it debuts in the market.

Why?

Due to the fact that there isn't a need for me personally yet for either the 8800/G92 series. By that time, (near nov/holiday season?), we'll start to get all the AAA titles and I'll actually want those new cards to push my screen to the max. As it is, my 7800gt is running just fine at 1920x1080 on older games as I see July-Oct as having no games that I need an uber graphics card for (with the exception that Crysis does still launch in Sept/Oct, then i take that back :p )

Either case, i'll either get a G92 if it outperforms 8800 series, or get an 8800 GTX for cheaper (hopefully) then as opposed to now.
 
What I'm hoping 92 is:

512MB Video Memory GDDR4
256-bit Memory Bus
128 SPs
600MHz Core Stock (650MHz factory OC)
1000MHz Memory Stock (1100MHz factory OC)
Physical size smaller then 8800GTS

And starting at $200 :D
 
YoNeX said:
For example, the 8800GTS has a peak performance of 345.6 GigaFlops (~345.6 Billion Floating Point Operations Per Second), while the 2900XT has peak performance of 475 GigaFlops. Clearly even an increase of 130 Gigaflops does not entirely show in the results.
Click to expand...
If I recall, NVIDIA said that G80 would hit about 500 gflop, so the one teraflop estimate is probably pretty high-ball. I wouldn't be shocked if G90/G92 could peak at about double G80 -- around 700 gflop, but I also wouldn't be shocked if it comes in a bit shy of that, given the necessary complexity of such a chip. If 700 is about right, I think we're looking at a minimum of a billion transistors (or a thousand million to some folk).
 
Isn't the G92 meant to be the midrange version? The G90 is meant to be the one with the 512bit bus and all the bells and whistles. Meant to come out in november as well (I have a terrible upgrade itch right now)
 
What I'm hoping 92 is:

512MB Video Memory GDDR4
256-bit Memory Bus
128 SPs
600MHz Core Stock (650MHz factory OC)
1000MHz Memory Stock (1100MHz factory OC)
Physical size smaller then 8800GTS

And starting at $200
Click to expand...

I hear ya brother.

Nvidia owes us a true midrange! No more 8600 garbage.
 
ryan_975 said:
The 65nm part is the size of the components that make up the GPU. Smaller compenents mean less voltage required to push electrons through the circuit (due to shorter distances). Lower voltage mean less resistance. Less resistance means less heat.
Click to expand...

No. The majority of power savings from die shrinks does come from the reduction in side of the gates (transistors), but not in the way you described.

The major consumer of power in chips is dynamic power (switching power), and this is proportional to the capacitive load of the transistors times the voltage squared. The larger the capacitors, the more charge it takes to make them switch. Dropping from say, 90nm to 65nm reduces the capacitive load by nearly half, and optimizing for the new process can help even more. It's not a function of "distance" so much as area.

The reduced size does affect the threshold voltage, which also allows for faster switching abd reduced source voltage, but it is typically not the biggest way you reduce power consumption.

That is of course assuming nothing else is changed inside the core. The Prescott P4's are the anti-case to what I just said, but that has a lot to do with the radical changes made to the P4's architecture (twice the cash, 11 more execution stages, etc) and the poor thermal design of the s478 platform.
Click to expand...

Actually, the huge power consumption spike with Prescott had to do with a combination of the poor architectural design (you mentioned) and Intel's failure with their first attempt at strained silicon (to reduce leakage losses).

Intel actually got their strained silicon working properly with their 65nm process. Cedarmill actually provides an excellent view into the benefits of simply shrinking the capacitors, because, aside from the process, it is identical to Prescott (same voltage, cache size, pipelines).

To get a image on what 65nm would look like: Imagine you have a 1 meter long stick. Then cut it up into a billion (literally) pieces. Grab 65 of them and put them back together. There you have a 65nm long peice of the stick. :)
Click to expand...

No, the gates have AREA, and they shrink in both dimensions. This is why the capacitance drops by half, because capacitance is also a function of surface area (90nm^2 is double that of 65nm^2).

If you want to see the difference, take a square piece of paper and cut off a third of the length in each dimension. The square you are left with is half the area.
 
ryan_975 said:
A flop is a FLoating point OPeration. A floating point number is any number from 0 to less than 1 (e.g 0.75 or 0.9939329353322).
Click to expand...

Right idea, wrong range.

A floating point number is named for the fact that the decimal point can "float," or move around. For example, using only 32 bits of data, a floating-point number has around 6 significant digits of accuracy. This means it can represent the values 32000.1, 89765000.0, 0.123456, and 0.0000987654 because they all have six significant digits.

Floating-point numbers are incredibly flexible, which is why they are used for 3D rendering. You can imagine that rendering the entire range of a 3D scene would use numbers both large and small.
 
I do think a lot of things are out of perspective. Namely DX10 performance and worth of the current cards. Currently DX10 is not rendering what some people would call earth shattering visual differences. Even though Crysis is certainly badass in DX10, the difference isn't "mind boggling!!!" like those who believed the hype thought it (DX10) would be. I think it will be another year at the earliest that DX10 starts to pull -far- away from DX9. There is simply no reason to own a low/mid-end DX10 card this early because the execution of the API is so low/early that you cannot get the performance to "tell enough difference".

As usual, Crysis will likely give us a better idea of how an 8600GTS card performs in DX10, same goes for the G92 (real mid-range performance likely) card and of course the upper 8800s. The game will definitely test the upcoming new mid and current high range cards for sure. I think it won't be as bad as most believe from the "proof" of 3-months-worth-of-effort DX10 patches. G90/1 hardware in general will handle the spec better I have no doubt. My guesses are a G90 will perform 20% over an 8800GTX OC'd/Ultra and the G91 will have a 50% performance increase over an 8800Ultra OC'd. I'm thinking this generational jump will be a little smaller, hopefully I'm wrong :p :D .
 
defaultluser said:
Right idea, wrong range.

A floating point number is named for the fact that the decimal point can "float," or move around. For example, using only 32 bits of data, a floating-point number has around 6 significant digits of accuracy. This means it can represent the values 32000.1, 89765000.0, 0.123456, and 0.0000987654 because they all have six significant digits.

Floating-point numbers are incredibly flexible, which is why they are used for 3D rendering. You can imagine that rendering the entire range of a 3D scene would use numbers both large and small.
Click to expand...

LIke I said. It was very late and my mind wasn't firing on all cylinders :D
 
You must log in or register to reply here.
Back
Top