View Full Version : What do you think the memory bus architecture of the NV30 will be?
Take your best shot at the memory configuration of the NV30
fooked if I know!! Although I'm pretty sure it wont be any of them 512 jobbies..
maybe it'll be 128, maybe it'll be 256. IF it's 128, they must have some tasty tricks up their sleeve to comabat the 9700, or they're gonna be toast!!
Comdex starts in about 5 days.. jeez I think this forum is gonna burst with the amount of "what size bus will NV30 have?" questions!! :D Its not that long to wait!!
StealthHawk
11-11-02, 09:13 PM
NV30 will be announced in one week, let's hope they give us specs and benchmarks and not just a "technology preview"
gemini1313
11-11-02, 10:43 PM
keep in mind the 9700 doesn't get any huge boost in performance from its 256bit bus unless u have max FSAA and AA turned on. also the 9700's fsaa and aa methods are hardly optmized to save bandwidth, so the increased amount speeds up its performance, Duh.
what is important is the speed of the chip without those items turned on, and then the speed with those on. also we know nv30 can compute a whole hell of a lot more than the 9700 with greater precision and longer shaders.
effective bandwidth really is the important factor, and u know nv30 has their bandwidth under control, as far as i know they have some new tech helping in that department.
Bigus Dickus
11-12-02, 01:24 AM
Originally posted by gemini1313
keep in mind the 9700 doesn't get any huge boost in performance from its 256bit bus unless u have max FSAA and AA turned on. also the 9700's fsaa and aa methods are hardly optmized to save bandwidth, so the increased amount speeds up its performance, Duh.The 9700 is highly optimized to save bandwidth during AA - both Z and color compression.
You are simply talking outta your ass here. Oh, care to explain the difference between "fsaa" and "aa?" :p
what is important is the speed of the chip without those items turned on, and then the speed with those on. also we know nv30 can compute a whole hell of a lot more than the 9700 with greater precision and longer shaders.Really? You know that? How many simultaneous operations can each chip issue and execute?
Oh, and the 9700 can caluculate a shader of the same length as the NV30.
Bigus Dickus
11-12-02, 01:27 AM
Just took a look at the poll again. What are some people thinking? Why would nVidia go backwards from a 4 way crossbar to a 2 way? It it has a 256 bit memory controller, there's no way in hell they would only put a 2x128 crossbar on there. Makes no sense.
Oh... and a 512 bit memory interface? :D :D :D
Originally posted by Bigus Dickus
The 9700 is [b]highly optimized to save bandwidth during AA - both Z and color compression.
You are simply talking outta your ass here. Oh, care to explain the difference between "fsaa" and "aa?" :p
Really? You know that? How many simultaneous operations can each chip issue and execute?
Oh, and the 9700 can caluculate a shader of the same length as the NV30.
That's true, but AFAIK, that's not correct for the Pixel Shader at all.
I'm not sure of the exact numbers, but I was under the impression from ATI papers that the R300 limit is 96 PS instructions while the NV30 limit is 1024 PS instructions.
Uttar
koneill
11-12-02, 02:25 AM
Originally posted by Bigus Dickus
Just took a look at the poll again. What are some people thinking? Why would nVidia go backwards from a 4 way crossbar to a 2 way? It it has a 256 bit memory controller, there's no way in hell they would only put a 2x128 crossbar on there. Makes no sense.
Oh... and a 512 bit memory interface? :D :D :D
NV30 will not have a 256-bit memory controller.
thcdru2k
11-12-02, 02:26 AM
is that a fact now?
SavagePaladin
11-12-02, 03:15 AM
<sneeze> this room is still giving me a cold, thats a fact <goes back to cleaning>
StealthHawk
11-12-02, 09:31 AM
Originally posted by gemini1313
keep in mind the 9700 doesn't get any huge boost in performance from its 256bit bus unless u have max FSAA and AA turned on. also the 9700's fsaa and aa methods are hardly optmized to save bandwidth, so the increased amount speeds up its performance, Duh.
and lets not forget how the R9700 dominates in games at higher resolutions too, even without FSAA.
and what is the point? that NV30 might be "good enough" with a 128bit bus? you just said yourself that the 256bit bus was crucial for super fast FSAA speeds. which leaves NV30 in a cold, dark place, as far as i'm concerned.
will that be the reality? will NV30 suck if it has a 128bit bus? i doubt it. R9700 has amazing tech, stop trying to belittle it. it's not just brute force.
PreservedSwine
11-12-02, 10:56 AM
Originally posted by Uttar
That's true, but AFAIK, that's not correct for the Pixel Shader at all.
I'm not sure of the exact numbers, but I was under the impression from ATI papers that the R300 limit is 96 PS instructions while the NV30 limit is 1024 PS instructions.
Uttar
That was incorrect, I think the R300 can excecute 256 instructions. However, the R300 has 32 texture instructions, and 64 ALU instructions each for scalar and vector. R300 can issue instructions from each of the instructions set each cycle, and consequently execute 3 instructions per cycle.
Anyone feel free to correct me on this....
jbirney
11-12-02, 10:57 AM
Originally posted by Uttar
That's true, but AFAIK, that's not correct for the Pixel Shader at all.
I'm not sure of the exact numbers, but I was under the impression from ATI papers that the R300 limit is 96 PS instructions while the NV30 limit is 1024 PS instructions.
Uttar
Uttar,
the R300 can loop back so any PS effect that can run on a NV30 can also run on a R300. Just that the R300 would have to do it in more passes. Kind of like DX8.1 and that whole PS1.3 and PS1.4 stuff
tazdevl
11-12-02, 11:32 AM
Read the David Kirk interview link on the front page. He pretty much comes out and says that it'll be 128bit.
We’ll move to 256bit when we feel that the cost and performance balance is right.
Originally posted by jbirney
Uttar,
the R300 can loop back so any PS effect that can run on a NV30 can also run on a R300. Just that the R300 would have to do it in more passes. Kind of like DX8.1 and that whole PS1.3 and PS1.4 stuff
Well, yeah. But it's obviously slower.
Uttar
Bigus Dickus
11-12-02, 12:47 PM
Originally posted by Uttar
I was under the impression from ATI papers that the R300 limit is 96 PS instructions while the NV30 limit is 1024 PS instructions.
Uttar
per pass IIRC.
Bigus Dickus
11-12-02, 12:50 PM
Originally posted by Uttar
Well, yeah. But it's obviously slower.
Uttar
Obviously? What is the number of cycles penalty for a multipass operation compared to the number of cycles needed to complete a 200 instruction shader program?
As the program length becomes greater, the penalty for multipassing becomes smaller. You're already talking about shader programs that might run a 1 fps a 320 x 480... is it really that terrible if it only runs at .88 fps?
Yeah, the penality is not too great. But I think you are kinda confused on what's those shaders use is.
You're already talking about shader programs that might run a 1 fps a 320 x 480... is it really that terrible if it only runs at .88 fps?
Such a shader will NEVER span the whole screen or an area as large as 320 x 480.
So... What's the use of those shaders? Their goals are to give additional details to SMALL parts of the screen. A good designer could guess where you're going to look at during a cutscene.
The same can be said for such a shader being used to increase skin quality ( as seen with the for-developer, publicly available, skin shader )
It'll NEVER be used for a 100x100 area. 3000 pixels max I'd guess. But that's just an estimate - it could be a little more or a little less.
But that brings an interesting point. What becomes important, thus, is the speed at which those shaders are switched, since that'll be the biggest performance bottleneck for amazing scenes where a lot of very small parts of the screen use specialized shaders.
In the Vertex Shader, Dynamic Branching fixes this. Static branching is of course also quite efficient, just a little less. We'll have to see the exact performance penality for nVidia's strange branching algorithm before judging I guess.
But... What fixes that in the Pixel Shader? Well, in PS3.0, there's branching. But neither the R300 or the NV30 supports that.
So, what did ATI & nVidia implement?
The R300 vs NV30 comparaison article at Beyond3D gives us the answer.
Another interesting thing about NV30 is that its PS instructions are stored in local video memory instead of chip interval. This method has two sides: it makes managing lots of fragment programs cheap though it also adds pressure to the already limited memory bandwidth.
Sounds like nVidia strategy is clear: The overall scene rendered with today's GPU is great, but attention to detail was forgotten before.
As for ATI, I don't know if they implemented such a thing. I'd guess the R300 got no such system, but maybe ATI simply didn't release the document. We'll see that soon I suppose.
Uttar
-=DVS=-
11-12-02, 03:05 PM
Originally posted by Uttar
That's true, but AFAIK, that's not correct for the Pixel Shader at all.
I'm not sure of the exact numbers, but I was under the impression from ATI papers that the R300 limit is 96 PS instructions while the NV30 limit is 1024 PS instructions.
Uttar
Its a fact that R300 does 256 instructions per pass but
NV30 in they papers it says 1024 instructions through Nvidia's CG - well what that means is they hideing the truth of actual single pass number ? , it will take few loops before they reach that big 1024 number :rolleyes:
its not a suprise all companys makeing exaggerations and truth bending :p
isay chill and wait for real hardware to see whats realy going on
Actually, I just checked the numbers, and it sounds like the 1024/pass number if for OpenGL. DX9 only supports 500/pass
And, AFAIK, there's no loop trick they could use here. There's no true branching/looping in the NV30/R300 PS.
Uttar
Bigus Dickus
11-12-02, 06:15 PM
Originally posted by Uttar
Yeah, the penality is not too great. But I think you are kinda confused on what's those shaders use is.
Such a shader will NEVER span the whole screen or an area as large as 320 x 480.
Uttar
Chicken and egg, huh? Don't you think if really amazing per-pixel lighting could be used on the entire map, it would? Why can't it? The speed of current shaders is just not up to the challenge.
That point aside, I still find it extremely unlikely that any game will use a shader program approaching 100 instructions during the lifetime of the R300 or NV30 (three years), so the argument is moot.
koneill
11-12-02, 10:35 PM
Originally posted by thcdru2k
is that a fact now?
Yes. That's a fact.
Originally posted by Bigus Dickus
Oh... and a 512 bit memory interface? :D :D :D
Yeah, I voted for that because I thought it was funny. :)
MuFu.
Originally posted by Bigus Dickus
Chicken and egg, huh? Don't you think if really amazing per-pixel lighting could be used on the entire map, it would? Why can't it? The speed of current shaders is just not up to the challenge.
That point aside, I still find it extremely unlikely that any game will use a shader program approaching 100 instructions during the lifetime of the R300 or NV30 (three years), so the argument is moot.
Yes. The current shaders aren't up to the challenge. However, it isn't very wise to use amazing per-pixel lighting on the entire map. It's a better idea to use amazing per-pixel lighting on a very specific areas then use the remaining power to have beautiful water and realistic-looking leather or other materials.
Well, actually, I think otherwise. I'm betting game with shaders which require about 100 instructions may be seen in about 18 months. Certainly not before, IMO.
But as I say so frequently, having longer shaders is a LOT easier than integrating a use for shaders in your program.
As for other DX9 features such as displacement mapping... Well, three years to be considered normal seems about right to me, with about 28 months before you even see a game using it right.
This is just my personal guess. It could be 100% wrong, but I really think that adding a small special effect using a 100 instruction shader isn't too hard if your engine is already made with PS & VS in mind.
Uttar
Mr Nvidia man, git off your @$$ and play with others outside ;) :D
You been listening to Carmack too much ;) :D
vBulletin® v3.7.1, Copyright ©2000-2012, Jelsoft Enterprises Ltd.