POLL: nVidia's Silence on GT300

[Keysplayr]

Originally posted by: HurleyBird

Originally posted by: Keysplayr
Can you guys define, "late to the party" for me please? I can understand if Nvidia announced a GT300 launch in June, July or August and it's still not here yet. But there have been no such announcements by them at all. You are only considering GT300 late because ATI is launching (pre-launching) R8xx soon.

I have no idea if Nvidia is "late" according to their own internal schedule or not, no non-insider does. That's different than being "late to the party" which just means that your product is out significantly later than your competitor's, and that's what counts. Keeping your schedule doesen't mean much when your shedule is way behind the competition anyway.

I'm merely making the observation that being significantly behind your competitor's schedule when you're releasing a new architecture has *never* been a good thing in the graphics world. It's not a prediction, just observation, and for all we know G300 will be the one to buck the trend, but it still gives AMD a window of oppurtunity to be the top dog.

Originally posted by: Keysplayr
One of the reasons (I think) AMD is able to release R8xx so quickly, is that R8xx (I think) is basically 2 R770's in one chip.

Anyway, A 4770 had a 128-bit memory interface. You can see each "core" in the schematice has 2 64-bit registers for the 256-bit total between both "cores". Also explains the doubling of everything. 800>1600 sps, 16 to 32 ROPs, 40 to 80 TMU's. This could also have been a contributing factor to the 4770 shortage for a while. Dedicating most of it's 40nm cores for Cypress, but I don't have any data to back this up. Grain of salt.

It sounds like you're saying that the RV740 shortage may have been due to AMD picking out the good ones to make Cypress dual core products out of? I'll attribute that to how late you're up posting :laugh:

Originally posted by: KeysplayrSo, at first glance, and if this schematic pic is accurate (unknown), it would seem that is exactly what AMD has done. Joined 2 800sp 4770's together, with other enhancements such as DX11 compliance, improved power circuitry, Eyefinity hardware.

All intercache bandwidth was effectively doubled, which is no small change, and I believe some caches were also doubled in size. There's a bunch of slides in the XS forums about this stuff. New scheduler too, and new aniso may or may not be hardware related. Just because something is structurally similar and looks the same on a simplified chart doesn?t mean that things are unchanged deeper down. Sure, G300 is almost certainly a bigger undertaking, but they?ve had enough time to work on it considering that all they?ve done since G80 is make it bigger with the slightest of architectural enhancements here and there.

Yes, we have established that everything has been doubled.

In bold ^
Yeah, Nvidia is able to generate an entirely new arch and deliver it only 2 months (November specualtion based on what rumors I've seen) after AMD's doubled up architecture with enhancements. If that doesn't impress you, then I'd like to know what does.

 

[dguy6789]

Why are you arguing the definition of "the party" and if someone is "late" or not? To me, it's when the first company releases a part, the next gen "party" has started. The second to come is late to the party if they don't launch simultaneously or near it. Being late isn't necessarily bad either. It's only bad if the new card doesn't offer enough advantages over the previously released card to justify the wait. There is of course no way at this point in time to tell when the GT300 will launch, how fast it will be, or how much it will cost. Everything is just speculation.

I don't see any problem with "doubling" an already successful architecture. I don't think anyone is trying to say that the R800 series isn't pretty much two RV770s or 740s or whatever with some tweaks. Neither having a new architecture or doubling an old necessarily correlates with winning or losing that particular generation. There are examples in history where new architectures have won and lost and double the previous gen architectures have won.

Edit: As far as release schedules, I personally don't care if a company releases a new architecture every 6 weeks even. If it's not better or not worth the money they charge for it, then it's not impressive no matter how fast they build it. Likewise, if a company does trump the previous champ, but takes forever to do it, that's hardly an accomplishment as well.

 

[Qbah]

Originally posted by: Keysplayr
Cookie, this is what I mean:

This As is

as opposed to

This As I would have expected

Keys... you do realize those two are actually the same (as far as parallelism goes - you omitted one section)? How can you not see that?

Each "SIMD Engine" is connected to the Ultra-Threaded Dispatch Processor and to the Global Data Share. You omitted the Global Data Share part (which means yours is totally different). They just drew it like that. They could do one column of the SIMD Engines, they did two, each Engine has the same connections though - so how is that "two cores in one package" ? You're totally off here.

 

[Qbah]

dp

 

[Keysplayr]

Originally posted by: dguy6789
Why are you arguing the definition of "the party" and if someone is "late" or not? To me, it's when the first company releases a part, the next gen "party" has started. The second to come is late to the party if they don't launch simultaneously or near it. Being late isn't necessarily bad either. It's only bad if the new card doesn't offer enough advantages over the previously released card to justify the wait. There is of course no way at this point in time to tell when the GT300 will launch, how fast it will be, or how expensive it will be. Everything is just speculation.

I don't see any problem with "doubling" an already successful architecture. I don't think anyone is trying to say that the R800 series isn't pretty much two RV770s or 740s or whatever with some tweaks. Neither having a new architecture or doubling an old necessarily correlates with winning or losing that particular generation. There are examples in history where new architectures have won and lost and double the previous gen architectures have won.

Edit: As far as release schedules, I personally don't care if a company releases a new architecture every 6 weeks even. If it's not better or not worth the money they charge for it, then it's not impressive no matter how fast they build it. Likewise, if a company does trump the previous champ, but takes forever to do it, that's hardly an accomplishment as well.

"Why are you arguing the definition of "the party" and if someone is "late" or not?"
Already explained.

"I don't see any problem with "doubling" an already successful architecture."
I don't either. Is there a reason you decided my observation suggested a problem?

" Likewise, if a company does trump the previous champ, but takes forever to do it, that's hardly an accomplishment as well."
As I have mentioned, a completely new architecture arriving shortly after it's competitors doubled and enhanced rehash (which isn't a bad thing, heading you off) of last gen tech, would be quite an accomplishment in most books even if it's not in yours.

 

[Keysplayr]

Originally posted by: Qbah

Originally posted by: Keysplayr
Cookie, this is what I mean:

This As is

as opposed to

This As I would have expected

Keys... you do realize those two are actually the same (as far as parallelism goes - you omitted one section)? How can you not see that?

Each "SIMD Engine" is connected to the Ultra-Threaded Dispatch Processor and to the Global Data Share. You omitted the Global Data Share part (which means yours is totally different). They just drew it like that. They could do one column of the SIMD Engines, they did two, each Engine has the same connections though - so how is that "two cores in one package" ? You're totally off here.

Yeah, mine is totally different. Are you kidding me? There is a data share BETWEEN the two cores in the original pic. How can you not see that?

Prove to me that they just "drew" it like that, and we will have a different conversation. I'm only going by what has been presented to us to date. Nothing more.

Hey, like I said, I'm open to other possibilities, and If actual x-ray die shots prove this schematic to be, "just the way they drew it" I'll have no problem retracting all my comments here.

 

[scooterlibby]

Originally posted by: taltamir


I am both surprised, unsurprised and saddened to see 11 people voted "Im an idiot and pick the first answer to see the results"

I, too, was in utter shock when I saw those results and had to wipe a tear due to the ensuing sadness. I had more confidence in anonymous internet poll voting, but my confidence was shaken to its core when I saw what had happened. Then, I realized I myself had voted for that option because it was pretty funny.

 

[HurleyBird]

Originally posted by: Keysplayr
Yeah, Nvidia is able to generate an entirely new arch and deliver it only 2 months (November specualtion based on what rumors I've seen) after AMD's doubled up architecture with enhancements. If that doesn't impress you, then I'd like to know what does.

If it does get released in November (and performs well) then I would be very impressed. I just don't think it's going to be released that early.

 

[dguy6789]

Originally posted by: Keysplayr

snip

My mistake. I read into it too much. I haven't had my coffee yet. :beer:

 

[Creig]

Originally posted by: Keysplayr
So basically that means their saying a 129.00 GTS250 could outpace a 5870 in a PhysX game. Well, duh!!! But did you interpret that to mean that you think Nvidia won't have anything soon to compete, so will offer GTS250's as a direct competitor? hehehehe.

No, that's not what I thought at all. hehehehe


Read Nvidia's statement again:

Because we support GPU-accelerated physics, our $129 card that?s shipping today is faster than their new RV870 (code name for new AMD chips) that sells for $399.

Where in there did they say the word "PhysX"? Hmmmmm? Nvidia is trying to trick the general public into believing that simply because their cards can do PhysX that they are generally faster across the board. Marketing dept misdirection at its finest.

Well, since they were obviously being deliberately misleading with that statement, why should we believe them when they say that they, "are absolutely confident that GT300 will win over Radeon HD 5870"?

They were probably just talking about PhysX again.

 

[Keysplayr]

Originally posted by: Creig

Originally posted by: Keysplayr
So basically that means their saying a 129.00 GTS250 could outpace a 5870 in a PhysX game. Well, duh!!! But did you interpret that to mean that you think Nvidia won't have anything soon to compete, so will offer GTS250's as a direct competitor? hehehehe.

No, that's not what I thought at all. hehehehe


Read Nvidia's statement again:

Because we support GPU-accelerated physics, our $129 card that?s shipping today is faster than their new RV870 (code name for new AMD chips) that sells for $399.

Where in there did they say the word "PhysX"? Hmmmmm? Nvidia is trying to trick the general public into believing that simply because their cards can do PhysX that they are generally faster across the board. Marketing dept misdirection at its finest.

Well, since they were obviously being deliberately misleading with that statement, why should we believe them when they say, "that they are absolutely confident that GT300 will win over Radeon HD 5870"?

They were probably just talking about PhysX again.

Ah, ok. If you say so.

 

[Creig]

Originally posted by: Keysplayr
Ah, ok. If you say so.

I say so.

 

[HurleyBird]

Creig has a point. When Nvidia marketing can make a statement that stupid --and I mean 11/10 on the stupid scale stupid-- it makes it hard to take anything they say completely seriously.

 

[Qbah]

Originally posted by: Keysplayr

Originally posted by: Qbah

Originally posted by: Keysplayr
Cookie, this is what I mean:

This As is

as opposed to

This As I would have expected

*snip*

Yeah, mine is totally different. Are you kidding me? There is a data share BETWEEN the two cores in the original pic. How can you not see that?

Prove to me that they just "drew" it like that, and we will have a different conversation. I'm only going by what has been presented to us to date. Nothing more.

Hey, like I said, I'm open to other possibilities, and If actual x-ray die shots prove this schematic to be, "just the way they drew it" I'll have no problem retracting all my comments here.

Check AMD's schematic again... Check yours... Still in denial or should I spell it for you?

AMD:
Each Local Data Cache is connected to the Ultra threaded Dispatch Processor
Each L1 Texture cache is connected to the Global Data Share channel and goes to the L2 cache - so each SIMD's L1 TC goes to L2
The Ultra-Threaded Dispatch Processor is also connected to the Shader Export

Where's the division? There isn't any. You can say whatever you like of course, that doesn't make it true - it makes it false. The picture clearly shows each SIMD engine being in parallel to every other SIMD Engine - you not being able to read the diagram and not seeing it doesn't make it any less true. They drew 20 SIMD Engines - means there's no more on the card itself right? Sheesh.

Your:
L1 Texture Cache goes to the Export Shader... why? There's no connection like that on the AMD picture. It's totally different. Meaning it's not the same. And there's no direct connection between the Ultra-Threaded part and the Export Shader part - how does it make the same? Your just shows "parallel for dummies, so everyone can see" (and changes the connections too!)- AMD's shows it's parallel too - you just don't see it.

I can draw you a picture if you still don't see it...

EDIT: Actually, there are 20 SIMD units, my bad. Doesn't change anything else I wrote.

 

[crisium]

Yeah, I have rarely heard such a display of arrogance from Nvidia as them actually saying that the GTS 250 is faster than the 5870. Especially since so few games support Physx.

 

[Cookie Monster]

Originally posted by: Keysplayr

Originally posted by: Qbah

Originally posted by: Keysplayr
Cookie, this is what I mean:

This As is

as opposed to

This As I would have expected

Keys... you do realize those two are actually the same (as far as parallelism goes - you omitted one section)? How can you not see that?

Each "SIMD Engine" is connected to the Ultra-Threaded Dispatch Processor and to the Global Data Share. You omitted the Global Data Share part (which means yours is totally different). They just drew it like that. They could do one column of the SIMD Engines, they did two, each Engine has the same connections though - so how is that "two cores in one package" ? You're totally off here.

Yeah, mine is totally different. Are you kidding me? There is a data share BETWEEN the two cores in the original pic. How can you not see that?

Prove to me that they just "drew" it like that, and we will have a different conversation. I'm only going by what has been presented to us to date. Nothing more.

Hey, like I said, I'm open to other possibilities, and If actual x-ray die shots prove this schematic to be, "just the way they drew it" I'll have no problem retracting all my comments here.

Think of it as SIMD cores sharing the L2 cache (e.g something like the L3 cache on nehalem being shared for all 4 cores). The word global data share could mean anything and in this situation, it looks like they've made a "unified" L2 cache that all the SIMD cores can have access to where as the RV770 did not have this.

R600 used a unified L2 texture cache
Notice how the L1 texture cache is connected to the L2 cache? If you look closely, it looks quite interesting because the RV870 seems to resemble some aspects of R600.

 

[Keysplayr]

Originally posted by: Qbah

Originally posted by: Keysplayr

Originally posted by: Qbah

Originally posted by: Keysplayr
Cookie, this is what I mean:

This As is

as opposed to

This As I would have expected

*snip*

Yeah, mine is totally different. Are you kidding me? There is a data share BETWEEN the two cores in the original pic. How can you not see that?

Prove to me that they just "drew" it like that, and we will have a different conversation. I'm only going by what has been presented to us to date. Nothing more.

Hey, like I said, I'm open to other possibilities, and If actual x-ray die shots prove this schematic to be, "just the way they drew it" I'll have no problem retracting all my comments here.

Check AMD's schematic again... Check yours... Still in denial or should I spell it for you?

AMD:
Each Local Data Cache is connected to the Ultra threaded Dispatch Processor
Each L1 Texture cache is connected to the Global Data Share channel and goes to the L2 cache - so each SIMD's L1 TC goes to L2
The Ultra-Threaded Dispatch Processor is also connected to the Shader Export

Where's the division? There isn't any. You can say whatever you like of course, that doesn't make it true - it makes it false. The picture clearly shows each SIMD engine being in parallel to every other SIMD Engine - you not being able to read the diagram and not seeing it doesn't make it any less true. They drew 20 SIMD Engines - means there's no more on the card itself right? Sheesh.

Your:
L1 Texture Cache goes to the Export Shader... why? There's no connection like that on the AMD picture. It's totally different. Meaning it's not the same. And there's no direct connection between the Ultra-Threaded part and the Export Shader part - how does it make the same? Your just shows "parallel for dummies, so everyone can see" (and changes the connections too!)- AMD's shows it's parallel too - you just don't see it.

I can draw you a picture if you still don't see it...

So why not link me to this AMD shot? That would have perhaps saved you a lot of typing and condescension. So, where is it?

 

[Janooo]

Memo to guys who think NV will release GT300 by November (within the next 10 weeks): keep dreaming.

How many discrete 40nm NV parts are out? They did not make a big 40nm GPU yet and suddenly 450-500mm2 will be ready? Keep dreaming.

 

[Keysplayr]

Qbah. Understand that whatever orientation the core "actually is, it makes no difference to me. Your denial comment indicates that you believe that I don't "want" the core to be one big core. I couldn't give two shits "how" it's configured, I'm just discussing what we see here. Do you have any sort of problem believing what I'm telling you here? Because I'm not going to discuss anything further with someone who flings the denial word at me as if I gave a shit on how it is configured. Dguy sort of did the same thing (although it was before his morning coffee) and interpreted my comments as me having a problem with the arch, which I do not.

 

[Qbah]

Originally posted by: Keysplayr
Qbah. Understand that whatever orientation the core "actually is, it makes no difference to me. Your denial comment indicates that you believe that I don't "want" the core to be one big core. I couldn't give two shits "how" it's configured, I'm just discussing what we see here. Do you have any sort of problem believing what I'm telling you here? Because I'm not going to discuss anything further with someone who flings the denial word at me as if I gave a shit on how it is configured. Dguy sort of did the same thing (although it was before his morning coffee) and interpreted my comments as me having a problem with the arch, which I do not.

Let me remind what you wrote:

One of the reasons (I think) AMD is able to release R8xx so quickly, is that R8xx (I think) is basically 2 R770's in one chip. Anyone who has seen the schematic photo for Cypress:

Cypress Architecture Schematic

You can see that there isn't one big core. It looks to me like they took 2 40nm 4770's (with the exception of 800sp instead of 640) and "glued" them.
"Glued" being a crude description, I'm sure there is more elegance to the design than that.

Anyone who saw that schematic saw that it's not two cores slapped together - only you saw it. All of the SIMD Engines are in parallel. AMD didn't slap two 800SP cores together in the new design. You said so because you mistakenly thought the drawing shows two cores slapped together but it doesn't - see the reason? If nobody said anything, you'd be saying in other topics about the architecture that it's just two cores together (based on what you think you saw on the drawings) - but that's not the case.

If you say the sun is dark just because you covered your eyes not to get blinded, doesn't mean it is anything other than bright. I would have a problem with you saying that because that is false and you'd have another point in your arguments later on (a false point). Even though I pointed clearly and without a shadow of doubt that the slide shows all SIMD Engines in parallel and not grouped into two distinct groups, you still claim that not to be the case. This is why I have a problem with your answers.

Most people get it - those that don't (even though they're wrong) need to eventually get it - especially you, who wants to participate in other discussions and have valid points in arguments.

 

[n7]

Don't know/care option.

I'm interested when the cards are out & for sale, same with AMD.

Anything prior to this is marketing-speak, which doesn't really excite me.

So no advance rumors/marketing blabber = i don't care.

 

[TheSlamma]

Is nVidia required to come out with a DX11 card the same day as ATi?

After looking at the benches in the 5800 thread nvidia still seems to be doing just fine with the 285 and 295's anyway. GOOD DX11 games won't be out for awhile & DX10 still seems to be sparse.

 

[geokilla]

Didn't want to search through 5 threads, but I find this funny. It's Hitler playing as NVIDIA's CEO as they talk about ATI's Evergreen release. http://www.youtube.com/watch?v=FR45ja_fNzU

On topic. I believe NVIDIA is behind schedule due to the at TSMC. Or maybe they want to see how the AMD GPUs will perform first before they release their own GT300.

 

[akugami]

Originally posted by: Keysplayr
Can you guys define, "late to the party" for me please? I can understand if Nvidia announced a GT300 launch in June, July or August and it's still not here yet. But there have been no such announcements by them at all. You are only considering GT300 late because ATI is launching (pre-launching) R8xx soon. And last time I checked, there is no rule saying both companies have to launch at the same time, and if one does, and the other doesn't, the other is late.

The only thing you can accurately construde from current events is, ATI is launching R8xx before Nvidia is launching GT300. This doesn't make GT300 late. When we get an official launch date for GT300, and that time comes and goes, THEN you can call GT300 late.

I'd call it (and this can apply to product lines from either ATI or nVidia) late if it is released well over a month later than whatever product it is competing against. For instance, if the GT300 was released 6-8 weeks after the Radeon 5xx0 series then I'd call it late.

A 4 week window of opportunity is great for ATI but I don't think it will make a significant dent unless ATI's new GPU's are better than nVidia's new GPU's. However, a 6-8 window (or more), would be pretty substantial. Keep in mind I'm not in any way shape or form saying the Gt300 will be that late. All of us really have no clue. Just illustrating an example. Another is how the X1800 was expected to release before whatever nVidia's new GPU was at the time. Because of bugs, the X1800 was released well after it's competing GPU. The X1800 wasn't a bad GPU but being late certainly did hurt ATI.

One of the reasons (I think) AMD is able to release R8xx so quickly, is that R8xx (I think) is basically 2 R770's in one chip. Anyone who has seen the schematic photo for Cypress:

Cypress Architecture Schematic

You can see that there isn't one big core. It looks to me like they took 2 40nm 4770's (with the exception of 800sp instead of 640) and "glued" them.
"Glued" being a crude description, I'm sure there is more elegance to the design than that.

Anyway, A 4770 had a 128-bit memory interface. You can see each "core" in the schematice has 2 64-bit registers for the 256-bit total between both "cores". Also explains the doubling of everything. 800>1600 sps, 16 to 32 ROPs, 40 to 80 TMU's. This could also have been a contributing factor to the 4770 shortage for a while. Dedicating most of it's 40nm cores for Cypress, but I don't have any data to back this up. Grain of salt.

I understand what you're saying but I don't think it will be that big of an issue if they actually did so. While it's not the most elegant or efficient solution it can be made to work. Intel essentially did the same thing with the Pentium D by slapping two P4's together. This staved off AMD's advance long enough for them to come up with a "true" dual core CPU. With the massive parallelism of GPU's I think "duct taping" two older (though updated) GPU cores might not be as bad as it sounds. There will be a performance hit for overhead but it should provide updated performance. Bottom line is performance, not elegance.

Also, I don't think we can make a claim that the GT300 is "new" while the Radeon 5xx0 series is "derivative" until we see more of the GT300 and Radeon 5xx0 and what each does differently or better than it's respective predecessor from a hardware standpoint. All of that will likely be dissected by January '10 but for now, we simply can't say. Too little information.

 

[MrK6]

My (very simple) assumption is the following:

AMD is on the verge of releasing a VERY impressive product. They completely improved everything about their 4xxx series and added new features - they are faster with better tech (40nm, etc.), better thermal profiles (27W idle!), better rendering (new AF), and new features (DX11, Eyefinity). On top of this, the high end card looks to compete head to head with nVIDIA's fastest card, which is dual GPU, using only a single GPU. A dual GPU version of this card is coming out a month later, indicating that AMD has some impressive performance solutions to offer this round.

There's this kind of information available and AMD is releasing within a week, and NVIDIA has done pretty much nothing to sell its (supposedly) coming line of cards. There's no "hey, here's a layout of our new architecture and why it's better than ATI's" or "check out this demo." There's been nothing but a few (idiotic) news articles stating things like "well, our cards are faster because we have PhysX" and "DX11 isn't that important." That's either being stupidly arrogant or desperation (and maybe both). Throw in some of these rumors that their having problems at the fab, and well, common sense lends one to believe everything isn't great on the greener side of the fence.

Anywho, that's how it looks to me, as an outsider. Maybe there will be more news or posts that will shed some light on this.