ATI responses about GeforceFX in firingsquad

[fbg1]

]The point is that by the time any games come out that would use such long shader routines, the actual routines would run at a snails paace even on the Nv30.. It will take a coupple full generational hardware upgrades to get to in game performance levels.

I've been wondering about that. It seems that NV30 has quite a lot of functionality and programmability, but lacks the bandwidth necessary to utilize 100% of it in real-time games. Both B3D and Extremetech's reviews suggest that as well.

However, could this functionality be put to use by movie studios like Pixar using NV30 to render animated movie frames? Such a task would not require real-time 30fps+. In fact, even 1fps might improve the cost-effectiveness of making a movie like Final Fantasy, as long as it was faster/cheaper than whatever tech has been used in the past. Further, given the choice b/t NV30 and R300, it appears that NV30 would be overwhelming favorite for this task given its full 128bits all the way through, and it's relatively massive shader programmability.

Everyone seems to assume that the main target audience of NV30 are gamers, but perhaps movie studios are just as important to Nvidia, hence the fully featured but bandwidth-limited NV30?

 

[RussSchultz]

In all honesty...not quite as tired of that as I am about seeing you post about it in every other thread.

Seriously, you are going to have to learn to tolerate it to an extent, because although it goes on to a degree here, there are at least valid points of view and good points and ideas being brought up in these threads as well, despite the fanboish tone at times.

The fanboism on B3D isn't anyhwere near that of sites like Rage3D or NVNews...but it's never going to completely disappear. You simply cannot "mod out" every simgle post that hints of fanboism...it would be counterproductive.

Ever wonder why I post this in every other thread? Because every other thread degenerates into (or starts out) a stupid tit-for-tat. And mostly its a SPECULATIVE stupid tit-for-tat from people who don't know what they're talking about in the first place and then extrapolate from their own ignorance. To top it all off, they then have the gall to start labeling each other as fanbois and denying that they're a pot or kettle also.

Hell, I really wouldn't even mind the product fanaticism if:
- they could stay on topic and not derail threads
- not repeat themselves in every thread
- only speak about what they know about.
- not use that damn eye roll emoticon.

But its not even fanboi-ism that I'm talking about. Its the pure immaturity that's growing on this forum. Quite honestly this forumis beginning to remind me of Jr. High.

There's threads that start out as interesting exchanges of information, but then get bogged down as somebody takes offense at how their prized product is being shown in light that isn't completely iridescent and drags in the same shite thats in 5 other threads to counteract whatever slight there was and we have another 10 page tit-for-tat that's just like the other five that were abandoned because they're exactly that. How many threads have monster fan jokes? How many threads are we talking about bus width? How many threads do we have speculating about when the RV350 will be out in february? How many threads have the jabs about when the NV30 will actually ship? Oh, every one?

You're right--you can't mod out every single post, but you can attempt to shame people into the accepted norm. If you're saying what we're heading toward is or should be the accepted norm, then have at it. You'll just have to put up with me whining, and the requisite people complaining about my whining.

 

[Bigus Dickus]

The framebuffer supports 128-bit floating point format, so no precision is lost.

Does the framebuffer support 128-bit floating point precision, or an off-screen buffer? If it's the framebuffer, is there a way to transfer 128-bit final rendered frames over the AGP bus before downconversion in the DAC's to a lower integer precision?

I know this is the eventual goal for rendering farms, but I've never seen anyone state that it can be done on the current R300 (or NV30) hardware.

 

[Dave Baumann]

Does the framebuffer support 128-bit floating point precision, or an off-screen buffer?

Technically, its an off screen buffer.

Friendly reminder to everyone: Search the forums!

 

[Joe DeFuria]

And mostly its a SPECULATIVE stupid tit-for-tat from people who don't know what they're talking about in the first place and then extrapolate from their own ignorance.

I disagree....there are certain instances of that...but in most cases it's people speculating in a reasonable sense. And in some cases the "nonsense" speculation from some, is put to good use by others who put out sensible arguments against it.

You're right--you can't mod out every single post, but you can attempt to shame people into the accepted norm.

Hehee...isn't that what flaming and fanoboi type posts try to do? "Humiliate" or "shame" the other side into changing their behavior (or outright leave?)

Seriously, I do get aggrivated and annyoed a bit myself for some of the very same reasons you mentioned.

How many threads are we talking about bus width?

Because that is one of the main distiguishing factors between R-300 and NV30 architecture? Not expected?

How many threads do we have speculating about when the RV350 will be out in february?

Will we EVER stop speculating on the next product?

How many threads have monster fan jokes?

How many threads have the jabs about when the NV30 will actually ship? Oh, every one?

Innoncent fun in most cases? Are you really bothered by these things?

Well, it seems to me you tend to be more vocal of these things where people are taking "jabs" and "jokes" specifically at nVidia's expense. What's this I found:

http://www.beyond3d.com/forum/viewtopic.php?p=48141&highlight=ati+drivers#48141

What are you doing take a jab at ATI's drivers for? Don't we get enough already of that here and "from the fanbois?" Do we need yet another joke about that?

There are lots of jabs done in good fun around here, and some that aren't. But I would suggest that unless you require everyone to follow a joke "in good fun" with an "explanation that this is just a joke" followed by a string of the SECOND most abused emoticon (winkey! winkey!) , then I would say that lowers the quality of this board as well.

 

[Bjorn]

but in most cases it's people speculating in a reasonable sense. And in some cases the "nonsense" speculation from some, is put to good use by others who put out sensible arguments against it.

I would agree with you if you said "in some cases it's people speculationg in a reasonable sense".

 

[BoddoZerg]

In thread "I can hardly wait for the NV30!":

Mostly because then I won't have to read page after endless page of speculation based on preconceived notions born in the world of fanboyism.

I suppose it won't end everything, because there's gits arguing about NV35 vs. R350 and beyond, but still...

Of course, there's also the endless whining about the endless speculation.... but I guess we can't have everything.

Ironic, isn't it?

 

[KimB]

And why exactly do you think this?
Despite all the facts (ATI has stated their current memory controller supports DDR2, ATI has shown a card running DDR2, etc...
Why would htey cut the bus width back? Do you even know what you are talking about?

To explain my position on this, I'll use a quick analogy:

Why do you think it is that the nForce doesn't usually gain much performance when you add in a second stick of RAM?

Obviously, it's because the CPU just can't access the memory any faster than the bandwidth of just one stick can provide. This is the same thing I see here. Unless the R300's memory controller was specifically made with higher-clock speed memories in mind (which seems doubtful, given the nearly-synchronous mem/core speeds currently, and ATI's synchronous clock speeds in the past), the R300 will gain little to no benefit from 256-bit memory at a higher clock speed than the core.

And let me reiterate:
Yes, the NV30 has a whole lot more fillrate in comparison to its memory bandwidth than the R300 does. Just because the "R300 has framebuffer compression too" doesn't mean it's going to be as efficient. A full analysis of the NV30 upon release will reveal which chip is more efficient. nVidia currently is claiming that it is, but yes, that does remain to be seen.

Said again, raw memory bandwidth means nothing. Effective memory bandwidth means everything (in all situations, not just a few...).

And I'm saying this because in recent history, despite a number of theoretical numbers that should have made the Radeon 8500 faster than the GeForce3/4, it wasn't.

 

[Doomtrooper]

Effective bandwidth is also limited by hardware limitations, no matter how exotic bandwidth saving features are included in the Nv30 if it only using a 128-bit bus it will not be able to compete in the FSAA and Anisotropic benchmarks vs 256 bit bus cards... IMO

A example is here from Ante P:

http://www.beyond3d.com/forum/viewtopic.php?t=3232&highlight=835

UT2003 Inferno 1024x768: Max FSAA Max Aniso:
GeForce 4 Ti 4600 (306/660): 8 fps
Radeon 9700 Pro (325/620): 75 fps

 

[Randell]

Effective bandwidth is also limited by hardware limitations, no matter how exotic bandwidth saving features are included in the Nv30 if it only using a 128-bit bus it will not be able to compete in the FSAA and Anisotropic benchmarks vs 256 bit bus cards... IMO

A example is here from Ante P:

http://www.beyond3d.com/forum/viewtopic.php?t=3232&highlight=835

UT2003 Inferno 1024x768: Max FSAA Max Aniso:
GeForce 4 Ti 4600 (306/660): 8 fps
Radeon 9700 Pro (325/620): 75 fps

OK do me a favour please DT to test this. Whats your Nature 1280x1024 4xAA/8x AF score. My score at 10x70 is 33fps, but my LCD cant do 1280x1024.

 

[Bambers]

Even the 9700 is memory bandwidth limited at very high res and FSAA isnt it? In 1600x1200 most of the benchs I've seen show the 9700 taking some reasonable hits from FSAA.

Would you still get those sized hits if it had 500MHz DDR II on a 256bit bus giving 32gb/s ?

 

[Joe DeFuria]

Chanloth,

I don't get your reasonsing. Can I try and clarify your position? I believe your stance is that ATI will come out with a board that uses 128 Bit DDR-II before they have one with a 256 bit bus and DDR-II. Correct?

Most people, including me, believe their first DDRII product will be 256 bit. For me, it's as simple as the reason being that that the more costly components typically get debuted in the higher priced products first, so there is less pressure on margins.

Having said that, you mentioned ATI's history of shipping nearly "synchronous" clocks. I don't think anyone here is suggesting that ATI will slap 400-500 MHZ DDRII on the already existing 325 Mhz R-300. I, for one, don't expect to see DDR-II on an ATI card until they can get the core clocks up to at least 400 Mhz on whatever part they pair it with.

...the R300 will gain little to no benefit from 256-bit memory at a higher clock speed than the core.

That's reasonable...but again, I don't think qnyone would argue against that. Your stance then, seems to be based on the premise that ATI won't raise it's core clocks from now until actually AFTER they introduce their first DDR-II products?

Again, if we are to be consistent with your "synchronous" observation, we know that ATI will NOT pair up DDR-II memory on a core that is less than 400 Mhz...because 400 Mhz is the floor for DDR-II.

So exactly what type of product (number of pipelines and clock rate), do you anticipate ATI will build for its first DDR-II part?

 

[OpenGL guy]

Does the framebuffer support 128-bit floating point precision, or an off-screen buffer? If it's the framebuffer, is there a way to transfer 128-bit final rendered frames over the AGP bus before downconversion in the DAC's to a lower integer precision?

I know this is the eventual goal for rendering farms, but I've never seen anyone state that it can be done on the current R300 (or NV30) hardware.

My understanding is that you can create a 128-bit renderable texture. Once you get the data into the texture, I assume you can lock it and read it back, but I don't know what DX9 has to say about that.

 

[Nappe1]

umm... is the DDR-II memory clocked to some rate?

theorically,
I mean why ATI could not take R300 / R350 core and clock it as high as possible with reasonable yelds and then downclock 400MHz specsed DDR-II to get syncronized clocks? Still with 256 Bit bus it would be faster than 128Bit bus variant. Of course, it would make it cost a bit more, but hey who cares about what it costs as long as it returns market leadership on high end...

 

[demalion]

Is anybody as tired of this as I am?

In all honesty...not quite as tired of that as I am about seeing you post about it in every other thread.

Actually, I think it was a pretty short and clear method of addressing his complaint myself. Very easy on the eyes, no long tirade, point made briefly. It is when the commentary on "fanboism" derail threads more than (or instead of) the "fanboism" itself that it strikes me as silly...so that's my last comment on it in this thread.

 

[Bigus Dickus]

umm... is the DDR-II memory clocked to some rate?

theorically,
I mean why ATI could not take R300 / R350 core and clock it as high as possible with reasonable yelds and then downclock 400MHz specsed DDR-II to get syncronized clocks? Still with 256 Bit bus it would be faster than 128Bit bus variant. Of course, it would make it cost a bit more, but hey who cares about what it costs as long as it returns market leadership on high end...

Yes, you can "downclock" DDR-II memory from, say, 400MHz to 350MHz. However, the performance will not be any greater than that which you get with 350MHz DDR-I memory, so there is no point in doing so.

Chalnoth's argument makes no sense. ATi already has a 128-bit 8 pipeline chip in the form of the 9500 Pro. They are not going to equip it with DDR-II memory because the cost/performance would put it in direct competition with an existing solution, the 9700 non-Pro. There is no logic in equipping a lower cost product with high cost high performance/speed DDR-II ram, thus crowding and confusing your product line.

It also makes no sense for ATi to reduce their flagship product from a 256-bit to a 128-bit bus. Suppose the R350 was actually 128-bit product. Unless they paired it with greater than 1.2GHz DDR-II memory, it would have less bandwidth than the current R300. Even with a higher core speed or improved architecture, it would probably only offer a marginal speed improvement, and would likely lose to AIB "factory overclocked" R300's running at 350/350 or better. With DDR-II memory stuck around 1 GHz for a while, the only way for ATi to increase performance (without major architectural changes) is to keep the 256-bit bus and increase memory speed. With DDR maxing out around 400MHz, this likely means DDR-II memory (or GDDR3, or whatever the case may be).

Now, the RV350 will probably be a lower cost replacement for the R300 to debut closer to the launch of the R400. By that time, perhaps 1 - 1.2 GHz DDR-II will be more available and cheaper. Whether it will be cheaper to keep the 256-bit bus and more complicated memory controller and PCB with cheap memory, or use the cheaper 128-bit interface with more expensive DDR-II, is an open question. It is the same debate we have had about the NV30 and R300, and is probably unanswerable without intimate knowledge of all manufacturing costs involved.

Perhaps the RV350 will have a 128-bit DDR-II memory bus, but I don't see how that is possible for the R350 if it is to match or exceed the R300's performance.

 

[demalion]

Chalnoth,

OK, correct me if I'm wrong...but wouldn't it be true if the the Athlon took advantage of the dual channel on nforce boards it WOULD show a performance increase? This wouldn't require a clock increase on the Athlon, but a change in its FSB architecture, and isn't that type of "change" already present on modern graphics cards? Seem like a strange analogy. But I'll try to transmute it to what makes more sense to me, and guess you are simply stating that there is a limit on what a GPU/CPU can read from a memory interface that is related to its own clock speed and memory access architecture. Let's see where I can go from there:

All recent Radeon cores that have asynchronous memory controllers have shown some (varying) benefit from increased memory clocks with the core speed remaining the same, even when the memory clock was faster than the core clock. The question is where does this stop? Is it where the actual data transfer, regardless of signaling protocol overhead, matches the maximum rate of the GPU? Where would that leave us? What was the efficiency of DDR again? (100/ by that efficiency) possible increase as an absolute upper theoretical bound then? Assuming that speculation is correct, won't that be an even higher increase possible for using DDR II, then?

And that is completely ignoring that the R300 core looks like it is viable in a higher (core) clocked product, and that we have no idea what the R350 has in terms of design goal.

Maybe the R350 is 128-bit DDR-II, or maybe the RV350 is, and so, maybe the 128-bit DDR-II will be first. I see this as quite possible, if it makes sense to produce a part like in the marketplace because I simply don't see why it couldn't be done this way. What I simply and completely can't wrap my mind around is why you keep saying it must be this way, because it couldn't be DDR-II plus a 256-bit bus.

Your reasoning on costs doesn't make sense at all (as far as I understand)...it doesn't have to be "1 GHz" -? Durned marketingspeak) DDR-II if it is on a 256-bit bus, so it doesn't have to even cost as much as the RAM on the nv30. I assume your reasoning is based on signalling problems over a 256-bit bus, but I'm unclear on why ATI would be tweaking a "Graphics" DDR-II derivative and not address that. Your reasoning on clock speed of the core doesn't make sense based on observed performance on R300 overclocking for example + the unknown nature of what R350 is designed to do. So I again I ask for clarification.

 

[Tagrineth]

And why exactly do you think this?
Despite all the facts (ATI has stated their current memory controller supports DDR2, ATI has shown a card running DDR2, etc...
Why would htey cut the bus width back? Do you even know what you are talking about?

To explain my position on this, I'll use a quick analogy:

Why do you think it is that the nForce doesn't usually gain much performance when you add in a second stick of RAM?

Obviously, it's because the CPU just can't access the memory any faster than the bandwidth of just one stick can provide. This is the same thing I see here. Unless the R300's memory controller was specifically made with higher-clock speed memories in mind (which seems doubtful, given the nearly-synchronous mem/core speeds currently, and ATI's synchronous clock speeds in the past), the R300 will gain little to no benefit from 256-bit memory at a higher clock speed than the core.

And let me reiterate:
Yes, the NV30 has a whole lot more fillrate in comparison to its memory bandwidth than the R300 does. Just because the "R300 has framebuffer compression too" doesn't mean it's going to be as efficient. A full analysis of the NV30 upon release will reveal which chip is more efficient. nVidia currently is claiming that it is, but yes, that does remain to be seen.

Said again, raw memory bandwidth means nothing. Effective memory bandwidth means everything (in all situations, not just a few...).

And I'm saying this because in recent history, despite a number of theoretical numbers that should have made the Radeon 8500 faster than the GeForce3/4, it wasn't.

Uh, the new increased FSB (333, 400) Athlons DO see a nice performance benefit using dual-channel. Whoops, there went that argument.

R300's memory controller doesn't have to be aware of higher-bandwidth solutions at all. It only technically has to be aware that memory exists. The driver (and maybe some firmware) does the optimisation for certain bandwidth amounts.

There's always room to improve anything. Increasing memory bandwidth will always result in a performance increase, at the very least in some areas. Likewise in some areas increasing the core clock speed will always introduce some kind of performance increase.

 

[Bigus Dickus]

I just had a stroke of genius... or ignorance. Not feeling too good today.

Anyway, perhaps the major architectural improvement of the R350 will be to increase internal bus to 1024 bits, allowing it to take full advantage (within limits of core/mem balance, of course) of higher clocked DDR-II memory.

A 400MHz R350 with 1 GHz + DDR-II memory and an internal 1024 bit bus would make a nice addition to my easter basket.

Oh well... perhaps all this damn cold and flu medicine is making me hallucinate. I think I'm going home. Screw work.

 

[Joe DeFuria]

1024 bit internal? I'd have to say I doubt that idea.

However, I think ATI could stand to improve its memory interface. Then R-300 is more or less the same carried over from the R-200, just "replicated." (R-200 = 2x64 bit, R-300=4x64 bit.)

I would like to see ATI go with a more granular "8 way cross bar", 8x32 bit for the 256 bit wide interface. (That would be like an extension of the current GeForce4 and GeForceFX interfaces).

Though having said that, the R-200/R-300 pipelines are probably more optimally set up for the larger 64 bit paths, which would likely mean a considerable redesign of the chip, which I don't feel is likely...