Is there something new on the horizon from nVidia?

[Transfurzz]

and transfurzz, as long as the discussion remains common to an architecture, FLOPs is a more than adequate description of relative performance for a given processor/memory system. Since all of these chips are G92, whether they have 112 or 128 shaders (alternatively, 506 or 624 gFLOPS at 600/1500 and 650/1700), predictions can easily be extrapolated. i'm not saying performance, or even FLOPS is proportional to the number of shaders on your chip. but what i am saying, is that if you have at least 55GB/s of memory bandwidth, the more shaders you have, the faster you'll be, and i was under the impression earlier the GX2 had 112x2 SPs.

what are you talking about. the gx2 does not have wham-bam performance over 2 GTs, let alone 2 GTS

So... you admit it ? Your operating on pure speculation ? You took a look at hardware specs, and then somehow concoct up some major theory about how the market will settle ?

How are you even sure that your FLOPs are the correct number ? Their two different pieces of hardware, how can you be so sure that the FPU is the same among those ? You'd figure of all the things they'd modify from one series to another, they would modify whatever aspect your going on around about now.

Why are you quoting memory speeds ? What does that have to do with FPU performance ?

FLOPs amount to nothing, it just measures the strength of your floating point unit and nothing more. There are numerous other things in the architechture you might not have taken into account. That's why it's only ever seen as a measure of number crunching ability. Not performance. Graphics cards aren't as one dimensional, nor as end-all, as you expect them to be. It's a TERRIBLE way to measure performance. The FPU isn't the only part that makes up the hardware, even a horrible piece of hardware can have an incredible FPU. In other words, FLOPs never equal to real world performance. That's why the emotion engine has more FLOPs when compared to the Xbox's Celeron.

Unfortunately it does have some wham-bam performance, apparently it's able to make Crysis very playable at 1680x1050. That's pretty damn good. When you start playing at resolutions at that, even a small increase is incredible. Notice it jumped from below 30 in average FPS to above 40 in average FPS ? That's a huge increase, you like to measure increases in percentage. Which to be honest, just diminishes the weightage of the entire aspect to nothing more than raw performance.

Furthermore, the stuff involved in the Call Of Juarez incident can't be wished away with just a single core revision. The article just says that Nvidia wasn't prepared for the changes that giving greater control to developers would bring. They didn't stick to DX10 specs very well as they knew their hardware simply didn't perform too well with shader-assisted MSAA resolves. Nvidia wanted them to work with it's own custom resolve, in other words they didn't heed the call of letting the industry have more control. ATi didn't complain about it at all, apparently Nvidia was caught off-guard.

How is this fixed ? Nvidia just has to make the cards beefier by quite a bit. The core revision wasn't much beefier, it just made the series run much cooler. The leap in performance needed to be larger. In other words, the Geforce 8 series actually wasn't the improvement in performance parts of the industry was looking for. The features are all there, and they run the majority of games well enough though. Consider that Crysis uses shader-assisted resolves quite a bit, that's how they get all that nice lighting. It's quite obvious Nvidia has put some thought into that area.

I hope you also do know that having more memory bandwidth has a positive impact on certain operations we all know and love. For example, AA and AF. Having more bandwidth can have a positive effect on either if the GPU is up to it. That's something you need to take into account too, especially at higher resolutions.

[iamsostupid]

How are you even sure that your FLOPs are the correct number ? Their two different pieces of hardware, how can you be so sure that the FPU is the same among those ? You'd figure of all the things they'd modify from one series to another, they would modify whatever aspect your going on around about now.

judging by your wording here, i can see you have no idea what i'm talking about. and you will have no idea until someone from anand, tom's of [H] puts it right in front of you: G92 is G92. nothing has changed.

how am i sure my FLOPS are the correct number? are you kidding me? nvidia has been pushing their G80/G92 architecture for years in GPGPU computing and they publish these numbers freely. they are on the website and wikipedia has made tables for 8 and 9 series cards. you will see memory throughput affects aggregate "theoretical shading FLOPS" quite a lot. the only person comparing these gFLOPS to CPUs, is you. I am comparing a common architecture with a common memory system. you are throwing in IBM and intel chips, which i'll say is a bit on the nose.


allow me to resurrect these old articles to go into the G80 architecture with you. read this page along with me:
http://techreport.com/articles.x/11211

The G80 has eight groups of 16 SPs, for a total of 128 stream processors. These aren't vertex or pixel shaders, but generalized floating-point processors capable of operating on vertices, pixels, or any manner of data.

there are a lot of handy diagrams in the review for you. you will see three schedulers called thread setup engines for vertex, geometry, and pixel. these run at the "core clock." what do they do? they prepare threads for massively multithreaded computation, performed by an array of FP32 MAD and FP32 MUL units clocked from 1.5-1.7GHz. so, can you explain to me again, because i'm slow, why FPU is irrelevant to GPU design/performance?


here you go. first page
http://enthusiast.hardocp.com/articl...50aHVzaWFzdA==

maybe you can believe them.
not surprisingly the G92 GTS SLI and 9800GX2 score within 1% of each other in all tests except COD4. In COD4, the GX2 cannot do TRSS AA at 2560x1600, so they ran it with no AA and it hardly beat SLI GT's. The SLI GTS did 4x TRSS AA just fine and averaged 62.8 fps.... whatever beef you are referring to, it remains to be spoken for.

Nvidia just has to make the cards beefier by quite a bit. The core revision wasn't much beefier, it just made the series run much cooler. The leap in performance needed to be larger. In other words, the Geforce 8 series actually wasn't the improvement in performance parts of the industry was looking for. The features are all there, and they run the majority of games well enough though. Consider that Crysis uses shader-assisted resolves quite a bit, that's how they get all that nice lighting. It's quite obvious Nvidia has put some thought into that area.

This paragraph insinuates that 9-series GPUs will put out less heat than the 8-series. This is incorrect in light of the fact that they use the exact same GPU. This paragraph also implies, in the most nondescript way possible, that the 9-series cards will be "beefier by quite a bit." What is this alleged "beef" you are talking about? I'm dying to know.

[OpStar]

Considering you can get 2 G92 GT's for under $400 and a GX2 is going to cost like $600... come on.

Right and the 2 GTs require a crappy SLi motherboard which adds additional cost. Not to mention traditional SLi is harder to setup and nowhere near as transparent. It also requires more power and outputs more heat. It also leaves less room for expansion.

The 9800GX2 is the only way to get that type of performance on an Intel chipset mobo.

You can use the data destroying 680/780i all you want, or you can upgrade to a 790i Ultra for $350 bucks, spending an extra $150 over the cost of the GX2.

But to sit here and say that better performance that requires less power, generates less heat and costs less than it did a year ago are bad things? Please.

Reviews have shown that the 9 series are highly overclockable. The GX2 is only going to get faster when it gets in more enthusiasts hands and not to mention driver revisions. The nvidia driver team has already greatly improved the GX2s performance between beta revisions to the release revision most everyone is testing on.

This is exactly what a videocard revision is supposed to be. Cheaper, less power, cooler, faster, more overclockable. While I agree maybe naming this series the 8900 series would've made more sense, they are still good cards.

You'll see on Tuesday when the 9800GTX reviews come out. It will be faster than the 8800GTX and cost $349 bucks. That is huge. Look how much cheaper performance has gotten in 18 months. Remember, the 8800GTX was $649 at launch.

[iamsostupid]

the 8800GTS is already faster than the GTX, except in a scenario where the memory system is unusually backed up. This still does not justify the $120+ price difference. At least for me.

Even if you were committed to an Intel chipset and wanted SLI performance, the GX2 is not justified at 649.

and i'll admit, when i read on xbit that the 9800GTX was going to be 349, I honestly said to myself "wow, that's nice." if they wanted any more for it, it would just be another farce. but please show me where a G92 9-series card has lower power consumption than a G92 8-series card...especially with these nondescript, extravagant clock speeds you're boasting.

This doesn't mean future nV flagships are all going to be $349 though. The only reason the card is so cheap is because it virtually brings nothing to the table that wasn't already on the table. Do you think GTX cards based on the nV GT200 are going to be introduced at $350? because they won't be... whenever that time comes.

The G80 made sense to be $650 at launch. It was the debut of a completely new architecture. It was a staggering leap in power and performance above the 7900GTX, and received no response from ATI (not for years). Since the G92 die shrink, nothing has changed ... so $349 is a fair price.

[Transfurzz]

Edit: Sorry Double Post

[Transfurzz]

Oh goodness, this is so simple it's laughable. FLOPs measure NOTHING but the performance of the floating point unit. It doesn't measure anything else. It is not a good descriptor of performance, simply because of the fact it takes nothing but the strength of the FPU in question. That alone as a benchmark makes it useless for videocards, it's merely a way to tell more knowing techies that we modified the FPU on this one. It'll be able to handle better calculation-intensive processes. For example, it doesn't measure how fast the information is loaded into the pipeline. GG with that. It's one of the reasons why the netburst architechture did so badly as when compared to the K8. Would you look at that ? Their rougly equivalent in terms of Gflops.

There is the supposed beef, because it's NOT the same bloody core. Look at it, it's doing a bloody extra 20 FPS more than an 8800GT SLI in Crysis with just 16 more stream processors. Doesn't that ring alarm bells to you ? That would mean that either video card is outperforming its 8800GT counterpart by a good 20FPS easily with just 8 stream processors. While a big contributing factor, I beg to differ, it's probably a lot more than extra shading power.

The power consumption ? I posted the article, it measures the rig it's in at an extra 133 Watts under load. The 8800GT by itself however, pulls an extra 74 watts under load. That's 148 watts for an SLi. The fact the 9800GX2 overclocks extremely well doesn't help your case either.

You talk as if all there is to the performance of a videocard, or any other piece of hardware with an FPU, is the strength of the FPU itself. You also use this as an excuse to compare two radically different products to each other, then accuse me of doing the same.

The fact is this, your waving around FLOPs, measuring value through shaders and not FPS, then accusing others of opinions which happen to be based in fact.

Oh wait ? Your first article doesn't support your views, so you run to another one ? Ohhhh, that's funny. HardOCP tests with very close attention paid to eyecandy. Look closely. Jericho ? Sheesh, hardly the industry benchmark. It uses a proprietary engine, you won't see that in any other game. The drivers certainly aren't built with that in mind. COD4 ? Even they admit it's more of a bug in the conclusion.

The fact is, the 9800GX2 offers good performance without an SLi motherboard and with less power consumption. It overclocks well too. It beats the 8800GT in SLi in many benchmarks, it also manages to match its power with just beta drivers. As this is a dual GPU card, the drivers are quite different from other cards.

[iamsostupid]

It *is* the same "bloody" core. the GX2 does not have 16 more stream processors than 2 GT's. It has 32 more stream processors. the GX2 has 256 aggregate shaders. Two 8800GT's have 224. This more than accounts for your missing 20 FPS.

If you look at the [H] review of the GX2 in crysis, you will see it is dead on with the G92 GTS SLI.

If nVIDIA had created a third-generation processor with 128 shaders (even if only for the specific application of the 9800GX2), not only would we have known about it ahead of time, but it would not be referred to as G92.

[Transfurzz]

It *is* the same "bloody" core. the GX2 does not have 16 more stream processors than 2 GT's. It has 32 more stream processors. the GX2 has 256 aggregate shaders. Two 8800GT's have 224. This more than accounts for your missing 20 FPS.

If you look at the [H] review of the GX2 in crysis, you will see it is dead on with the G92 GTS SLI.

If nVIDIA had created a third-generation processor with 128 shaders (even if only for the specific application of the 9800GX2), not only would we have known about it ahead of time, but it would not be referred to as G92.

No it isn't. Regarding Crysis, those are the highest playable settings available. According to the Bjorn3d review, HardOCP is inaccurate. As the true deficiencies of either video card are more accurately measured at a lower resolution. Namely, 1680x1050. Crysis pushes the envelope, the settings they are using makes every video card struggle with it. How can you tell there is a performance leader then without backing down on the settings.

In fact, if you scroll down. The article clearly reveals that it can run AF much more comftably than the 8800GT can at 1920x1200. Testing for performance and playability are two different things.

HardOCP perfers to do the latter, as it thinks performance benchmarking is not reflective of real world performance. Actually, I couldn't care less. Their graphs and charts make little sense to me, they don't present data properly. I can understand why you made that mistake there, I think it's stupid of them to put that thing regarding AF all the way down.

It is not a carbon copy of the 8800GT core. It's based on the same design and stems from the same design. But it is not the 8800GT core. You'd do well to remember that. It has never worked that way. A G92 core can have 8 stream processors too. It just has to be based on the same design. Heck, the 8800GTS G92, which you have conveniently forgotten is a testament to the fact. A card can be of the same core design, but very different.

Oh come on, it's doing all this on a lower core clock than the GTS G92. Which is the true SLi combo it runs neck in neck with. You'll have to at least say they beefed it up, it's based on the same core design, so what ?

Heh, you make fun of me for using wikipedia too.

[James]

I think the point that you guys seem to be missing...

9800GX2 = G92 GTS cores, not GT cores.

G92 GTS = 128 stream processors at a default 650MHz.

G92GTS(x2) = 256 stream processors at a slightly lower 600MHz. Considering they are sharing a HSF unit for the two of them, I personally think that's pretty impressive.

Maybe not $600 impressive, but impressive nonetheless.

[iamsostupid]

lol, we know. it's just that he is trying to convince me that there is something MORE to the GX2. Something that he cannot tell me, because he doesn't know.

he thinks that since the GX2 is faster in a few specific scenarios (despite being clocked lower than the GTS) that this is evidence enough of some kind of additional, unmentioned hardware on the card. he has thus far referred to it as "beef," but I have my doubts.

I do not believe nVIDIA would ramp production of a new, updated GPU, so late in the G92 product cycle, with GT200 right around the corner, for such a premium, low-volume product (GX2). Granted, the 9600GT and 9500-series cards do have new GPUs (G94 and G96... 65nm die shrink and updated texture compression over the G84 and G86), the G92 processor already supports these new compression algorithms, and is already on 65nm.

Has nVIDIA designed a new card for add-in board makers? Yes, they have.

Has nVIDIA asked UMC or TSMC to build GPUs with 128 shaders OTHER than G92? No, they have not.

[James]

I'm just hoping their high-end and 2nd to high end of next gen cards have 512-bit or 384-bit buses again.

There are definite points (AA enabled in resolutions 1920x1200 or higher) where the extra memory bandwidth is vital. As I plan on owning a 24" LCD widescreen with my tax rebate (it's a maybe at this point) that is something that I look forward to.

[iamsostupid]

no 9-series card will have greater-than-256-bit-wide memory. sry.

[OpStar]

512 bit bus would've cost a fortune, which is probably why they didn't do it. Not to mention the fact that the 2900XT had it and sucked anyway.

[James]

True, but nVidia actually has the horsepower to take advantage of the bandwidth it seems, at least up to the 384-bit bus on the 8800GTX/Ultra

[vader545]

Just FYI: According to the genus book of world records the world’s most powerful computer runs at 8.5 petaFLOPS.