So Gubbi it would be okay for me to create a 3ghz chip that gives the same performance as a 1ghz chip but that is okay because i designed it to clock that high and to get that speed ? Hell no. Overclocking my radeon 9700pro with stock cooling i can get it up to 370mhz. If i took a flowfx and put it on there i'm sure i can get even more out of it. The geforce fx is in the same spot as the radeon 8500. Same speed as the geforce 3 ti 500 and offered enhanced features over the other cards ... Yet it was to little to late cause it got pounced by the Geforce 4s. That is what will happen with The geforce fx. Unless they got the low k working this chip reached its thresh hold . Lets hope the nv35 does a whole lot of things right and the r350 does alot of things wrong. And lets hope nvidia can get decent drivers for the card ...
nVidia's Jen-Hsun Huang speaks out on X-Box 2 possibility.
- Thread starter Brimstone
- Start date
jvd said:
Do you have any idea how bad of an argument you're making? Clock speed is dependent upon many factors well beyond you're Overlocking Slider that you drag over in your nice GUI menu.
Designing to meet a particular clock speed is just as much of an architectural challange and design achomplishment as an 8 or 16 pipeline architecture.
Infact, preformance can be seen as a delicate balance and result from a equation that contains not just sheer gate counts, but also clock speed and architectural effeciency.
Steve Dave Part Deux
Newcomer
I thought it was evident by now that the GFFX was not originally designed to run at 500Mhz. GPU speed is more tied to manufacturing process than pipeline length. In fact, the nature of 3D graphics processing lends itself much better to increased parallelism and that fact has driven GPU design since its' incarnation. Besides, the pipelines are nearly identical featurewise(they were both built with the same target spec - DX9). I think such a comparison is deserving of at least an Golden Delicious - Macintosh rating.
Vince his is just as stupid as mine. He tells me to go look for a radeon 9700 pro clocked at 500mhz . A non overclocked 9700pro. Of course there isn't one. Its not because it can't be done. Its because ati has no need to do it. They have equaled nvidia on an older micron and a slower clock speed. To me that means the r300 is clock for clock faster and over all = to the nv30. The greater feat is the r300 . The nv30 is late it shouldn't even be compared to the 9700pro as its not on the shelf yet. Who knows it may get on the shelves after the r350.
BenSkywalker
Regular
JVD-
I'm not talking about it in a PC add in board capacity on that end, I'm talking about the financial end for console useage. And on ATi having a faster board out soon, almost everyone assumed the same about nVidia when the R300 hit.
The NV30 is forced to run in FP32 using the current ARB path as it doesn't allow using FP16 if the board supports higher precission. This is something that is being rectified in the final build. As far as actually getting the NV30 up to speed, how long do you think Carmack has actually had a NV30? We know he has had a R300 since last May at least seeing he showed D3 at E3 on one.
Tag-
Why would they even claim a noise level that low unless they had something that could deliver? At 20db it would still be far less noise then any other top tier card, I would say that certainly have something that removes the noise issue for FX's.
If you notice, I've been listing 4x AA numbers when possible, and in past conversations we've had I've been quite adamant in stating that I expect 4x AA @1080i to be the norm next gen. I bring this up as they would actually be rendering @3840x1080 vs the PC running @2054x1576-2560x1920/2048 using the resolutions I was using to compare. The actual resolution of 1080i is 1920x540.
You mean obsolete pixel shaders(certainly not PS 2.0/3.0)
PS 1.4 native test, I am looking forward to seeing the next revision of 3DMark to see what current pixel shader performance is at.
My assumption would be that they are forced in to FP32 by default(although it is possible that this could be altered by a future driver revision), another reason to look forward to a DX9 level shader bench so they can specify the level of precission used. Pixar uses a FP16 color format for their movies, I don't think we are quite at the point where we need higher precission in real time quite yet
I've seen reports saynig that the .15u R350 will be the only one out in the near future too, those various bits of information floating around are what they are, rumors. It was rumored that the NV30 taped out long before it did, when I see something definite from ATi I'll believe it a lot more then what I do now.
By allowing vendors to low level access to their hardware they could assure the backwards compatibility will only work with their chips.
Ever read any of the patent threads that go on here? It regularly takes years between the time a patent is applied for and when it is granted in the tech sector.
JVD-
Hell yes. Wait until Hammer core chips show up this year, they will likely be running with P4s running at 60%-70% higher clock speed in performance terms.
Compare the Voodoo3 to the GeForce1. The GF1 was clocked at 120MHZ while the V3 had versions clocked at 166 yet the GF1 throttled it completely. Even in the CPU space clock speed is a distant second in any real world terms outside of marketing, in the GPU space it has never really meant much at all when comparing chips of different architectures.
Why hope ATi stumbles? It is illogical to want a company to do poorly particularly when we are looking at two companies that for all practical purposes are within spitting distance of each other and are easily strong enough to keep pressure on the other. I wouldn't worry about the NV30 drivers, Carmack has already stated their OGL offerings are better then the R300's, a pre production preview piece of hardware versus one that has been out half a year. It is possible that the NV30 will ship with major problems with drivers, but based on nVidia's history for the last several years I find that extremely unlikely.
On the flip side, the R300 has quite a bit more bandwith so the NV30 is much more efficient then the R300. We don't know that either of these are true until we can see some good OCing numbers to get them comparable to each other in terms of core clock and mem bandwith.
I'm not talking about it in a PC add in board capacity on that end, I'm talking about the financial end for console useage. And on ATi having a faster board out soon, almost everyone assumed the same about nVidia when the R300 hit.
The NV30 is forced to run in FP32 using the current ARB path as it doesn't allow using FP16 if the board supports higher precission. This is something that is being rectified in the final build. As far as actually getting the NV30 up to speed, how long do you think Carmack has actually had a NV30? We know he has had a R300 since last May at least seeing he showed D3 at E3 on one.
Tag-
Why would they even claim a noise level that low unless they had something that could deliver? At 20db it would still be far less noise then any other top tier card, I would say that certainly have something that removes the noise issue for FX's.
If you notice, I've been listing 4x AA numbers when possible, and in past conversations we've had I've been quite adamant in stating that I expect 4x AA @1080i to be the norm next gen. I bring this up as they would actually be rendering @3840x1080 vs the PC running @2054x1576-2560x1920/2048 using the resolutions I was using to compare. The actual resolution of 1080i is 1920x540.
You mean obsolete pixel shaders(certainly not PS 2.0/3.0)
PS 1.4 native test, I am looking forward to seeing the next revision of 3DMark to see what current pixel shader performance is at.My assumption would be that they are forced in to FP32 by default(although it is possible that this could be altered by a future driver revision), another reason to look forward to a DX9 level shader bench so they can specify the level of precission used. Pixar uses a FP16 color format for their movies, I don't think we are quite at the point where we need higher precission in real time quite yet
I've seen reports saynig that the .15u R350 will be the only one out in the near future too, those various bits of information floating around are what they are, rumors. It was rumored that the NV30 taped out long before it did, when I see something definite from ATi I'll believe it a lot more then what I do now.
By allowing vendors to low level access to their hardware they could assure the backwards compatibility will only work with their chips.
Ever read any of the patent threads that go on here? It regularly takes years between the time a patent is applied for and when it is granted in the tech sector.
JVD-
Hell yes. Wait until Hammer core chips show up this year, they will likely be running with P4s running at 60%-70% higher clock speed in performance terms.
Compare the Voodoo3 to the GeForce1. The GF1 was clocked at 120MHZ while the V3 had versions clocked at 166 yet the GF1 throttled it completely. Even in the CPU space clock speed is a distant second in any real world terms outside of marketing, in the GPU space it has never really meant much at all when comparing chips of different architectures.
Why hope ATi stumbles? It is illogical to want a company to do poorly particularly when we are looking at two companies that for all practical purposes are within spitting distance of each other and are easily strong enough to keep pressure on the other. I wouldn't worry about the NV30 drivers, Carmack has already stated their OGL offerings are better then the R300's, a pre production preview piece of hardware versus one that has been out half a year. It is possible that the NV30 will ship with major problems with drivers, but based on nVidia's history for the last several years I find that extremely unlikely.
On the flip side, the R300 has quite a bit more bandwith so the NV30 is much more efficient then the R300. We don't know that either of these are true until we can see some good OCing numbers to get them comparable to each other in terms of core clock and mem bandwith.
Okay ben i see where your coming from in the clock speed area.
I also didn't mean that i want to see ati fail. I meant it as in i want them to even themselves out. 6 months ago ati did alot of things right and nvidia did alot of things wrong. Now i hope it evens it self out so that the nv40 and r400 are close in speed.
I'd also like to see a comparison of the r300 and nv30 at the same clock speed cause i really don't by your mem bandwidth idea. Sure ati used a 256bit bus. But they used 310mhz ram instead of 500mhz ram. But its stupid to debate anything. \
Your statements on the ARB not using fp16 on the nv30 is something i will have to remember and throw back at you when its fixed and the radeon is using fp24 and the nv30 is using fp16. Also he never says what fp the nv30 path is using it could even be using interger.
I also didn't mean that i want to see ati fail. I meant it as in i want them to even themselves out. 6 months ago ati did alot of things right and nvidia did alot of things wrong. Now i hope it evens it self out so that the nv40 and r400 are close in speed.
I'd also like to see a comparison of the r300 and nv30 at the same clock speed cause i really don't by your mem bandwidth idea. Sure ati used a 256bit bus. But they used 310mhz ram instead of 500mhz ram. But its stupid to debate anything. \
Your statements on the ARB not using fp16 on the nv30 is something i will have to remember and throw back at you when its fixed and the radeon is using fp24 and the nv30 is using fp16. Also he never says what fp the nv30 path is using it could even be using interger.
BenSkywalker
Regular
JVD-
Pixar is currently using FP16 for their movies, CGI quality for current rasterizers I'll take(not that they are there, but not due to lack of color precission). Carmack has been asking for 64bit color for a while, noone I've seen has been wanting 96bit or 128bit in the near future. Show me somewhere that is makes a difference in quality between FP16 and the higher precission levels(beyond FP16) and I'll gladly accept that there is a good reason for higher levels of precission, I haven't seen one yet.
That would have been something he would have mentioned.
Pixar is currently using FP16 for their movies, CGI quality for current rasterizers I'll take(not that they are there, but not due to lack of color precission). Carmack has been asking for 64bit color for a while, noone I've seen has been wanting 96bit or 128bit in the near future. Show me somewhere that is makes a difference in quality between FP16 and the higher precission levels(beyond FP16) and I'll gladly accept that there is a good reason for higher levels of precission, I haven't seen one yet.
That would have been something he would have mentioned.
Ben i'm not saying that it needs it . I'm saying that the radeon will be donig it in that format. As he said you can only use the best quality on the radeon. So that would make a comparison moot.
See i find it odd that he didn't say anything about what format the nv30 path uses. ITs whats not said that is where you should look.
See i find it odd that he didn't say anything about what format the nv30 path uses. ITs whats not said that is where you should look.
Real funny.
I was under distinct impression that 'actual' resolution is what we see on screen. If it's refreshed in fields, tiles, or concentric circles, what difference does it make?
And even the "internal resolution" thing becomes a tacky issue when refresh rates go to 100 and over.
Do you really want to double your framerate just so you can keep rendering in fields (which would mean you render the same amount of pixels in the end anyhow) when half the fps is already at 60?
jvd said:
I have problems with this statement - as increasing the clock frequency is basically Free from a manufacturing point of view if the yeilds are there. ATI is the undisputed underdog, they need to prove themselves - I can hardly believe they're selling all their R300s substantially underclocked just because they want only parity with nVidia - wait, they barely have that with the FX.
It just doesn't add up. Not only is Common Sense a good book, but a usefull tool in life.
BenSkywalker
Regular
Faf-
If you look at it like that then you are talking about half the framerate, that does make a difference when comparing it to a progressive display.
When TVs hit refresh rates over 100 that would make for a good point
Absolutely, to reduce input latency and accuracy. If all you care about is graphics that's one thing, I want some gameplay to go along with it.
JVD
That is a design shortcoming of the R300, not offering the mode developers want to use. You could say the same thing when they desire FP24 and the FX lacks it, as soon as that happens. The comparison isn't moot if they both output the same quality.
Why? We know that the ARB path forces the NV30 to use FP32 under the prerelease build if it uses FP at all. We know Carmack has been pushing for FP color for years. Why would he revert to integer color when he's been asking for FP for years? Do you think he's some kind of raging nV fanboy?
If you look at it like that then you are talking about half the framerate, that does make a difference when comparing it to a progressive display.
When TVs hit refresh rates over 100 that would make for a good point
Absolutely, to reduce input latency and accuracy. If all you care about is graphics that's one thing, I want some gameplay to go along with it.
JVD
That is a design shortcoming of the R300, not offering the mode developers want to use. You could say the same thing when they desire FP24 and the FX lacks it, as soon as that happens. The comparison isn't moot if they both output the same quality.
Why? We know that the ARB path forces the NV30 to use FP32 under the prerelease build if it uses FP at all. We know Carmack has been pushing for FP color for years. Why would he revert to integer color when he's been asking for FP for years? Do you think he's some kind of raging nV fanboy?
Actually, no.
Overlapping of fields from different frames will happen at half the framerate too - less often, yes, but with bigger gap between frames, more noticeable.
As with normal progressive display, there's only one way to actually fix that - increase refresh rate until the eye no longer notices it.
From what I remember some PAL tvs already do
not sure how it's elsewhere though.That I agree with, but I would prefer to have stable framerate first, reduced latency second
I find that hard to believe since the Renderman spec simply specifies float (and most Renderman renderers that I've used or built use 32-bit floats or depending on the CPU and compiler will use doubles internally).
[qoute]CGI quality for current rasterizers I'll take(not that they are there, but not due to lack of color precission). Carmack has been asking for 64bit color for a while, noone I've seen has been wanting 96bit or 128bit in the near future. [/quote]
Actually depending on the renderer, some commercial renderers are computing 128, 192, 256-bits. (read: 32 to 64 bit per component in either float and/or int) Which is probably why (aside from zero platform optimization) are usually agonizingly slow...
Aside from that 1080i signals actually contain 1125 scan-lines
Heh, a lot more common that you think! My first two Hi-Vision sets supported 100-120Hz refresh rates (dunno about my current set though, it's an American ASTC set and I haven't found any specs on it)...
Actually I think it's Nvidia's drivers that forces that not ARB_fragment_program... The ARB extension *does* support precision hints for hardware that supports more than one level of precision (dunno if Carmack is taking advantage of that or not though).
BenSkywalker
Regular
Archie-
PRRenderman used FP16 for every version I've ever seen, don't know about the generic Renderman renderer.
Which ones? Last version of Mental Ray I used ran FP16, same with 3DSM and Lightscape.
Do you know of any US sets that support 100/120Hz? I've never seen or heard of one.
According to the latest ARB notes they just added this in.
Faf-
No matter how you look at it 1080i is only displaying 1920x540 pixels @60Hz. Given how TV standards advance in the US that will likely be the highest end solution for decades.
Of course, but rendering half the pixels up front will certainly aid on that end
PRRenderman used FP16 for every version I've ever seen, don't know about the generic Renderman renderer.
Which ones? Last version of Mental Ray I used ran FP16, same with 3DSM and Lightscape.
Do you know of any US sets that support 100/120Hz? I've never seen or heard of one.
According to the latest ARB notes they just added this in.
Faf-
No matter how you look at it 1080i is only displaying 1920x540 pixels @60Hz. Given how TV standards advance in the US that will likely be the highest end solution for decades.
Of course, but rendering half the pixels up front will certainly aid on that end
FP16
Hi Ben,
3DS max ( and earlier versions right back to 3DS4 ) has always used FP32. The Renderman language only specifies float, and the RIB handles single and double precision IEEE formats.
As it is more difficult to implement FP16 in software than it is to actually just use hardware FP32 on every machine going back to the 68020/68881 I would be very surprised to find a single piece of software that used it.
On dedicated HW strange formats such as FP16 or FP24 make sense as they can save gates and decrease latencies, and FP16 for video cards is much better than the pure integer calculations supported before.
Hi Ben,
3DS max ( and earlier versions right back to 3DS4 ) has always used FP32. The Renderman language only specifies float, and the RIB handles single and double precision IEEE formats.
As it is more difficult to implement FP16 in software than it is to actually just use hardware FP32 on every machine going back to the 68020/68881 I would be very surprised to find a single piece of software that used it.
On dedicated HW strange formats such as FP16 or FP24 make sense as they can save gates and decrease latencies, and FP16 for video cards is much better than the pure integer calculations supported before.
BenSkywalker
Regular
Hey Crazyace-
All the information I can find indicates that 3DSM uses 64bit color(16bit per channel), along with Mental Ray, Lightworks and PRRenderman.
All the information I can find indicates that 3DSM uses 64bit color(16bit per channel), along with Mental Ray, Lightworks and PRRenderman.
Not NEARLY as agonizingly slow as they would be if they ran FP16 emulation though... 8)Archie said:
Ben,
as long as the renderer is running off CPU, I can't even imagine it running FP16, heck I don't recall ever seeing a general purpose FPU with that format. As Ace said, you would have to emulate it, which would among other things make it horribly slow.
I don't know about ports of various renderers to more exotic hardware but x86 versions are running FP32/FP64, I'm sure of it.
In which case I again put faith into seeing more sets capable of higher refresh rates in that resolution too. Assuming full frame rendering being the norm next gen, those kind of sets would effectively also make the argument about number of pixels moot.
R300 offers all the formats that developers want to use - its inputs and outputs are actually more flexible then NV30 in some respects. What goes on internally should be of little concern to developers as long as the output is correct. Seeing as you are arguing that 'FP16 is enough' then I assume you accept that FP24 will be suficient as well.
1080i do have higher resolution than 540 lines. Not as high as 1080p (due to the flicker filter), but still higher than 540p. If there was no advantage to running 1080i over 540p, then they wouldn’t use 1080i. Just compare any game running 480i (like most games do) with one running 240p (like Ico). Whatever you say, it is obvious that the 480i game has higher resolution than the 240p one.BenSkywalker said:
