PDA

View Full Version : Who Has It All Wrong.. Me Or Him


Drumphil
10-02-03, 12:28 PM
QUOTE ME:

Show me anywhere where I said that fp32 is a standard.. whats that got to do with it... If you run in fp24 mode, the NVIDIA cards DO render it in fp32, because they can't do fp24 (which is one of the BIG reasons they are slower than ATI card when running at dx9 full precision (fp24) ). What exactly is it about fp24 and fp32 that you think I got wrong?? I don't understand your complaint

to quote you: "And yes NV 35 can and will run pure FP32 if any aplication comes out in future........."
.. but don't you say that the drivers will replace all shaders with mixed precision ones??.. Why would that change once fp32 comes in.. How will that be different to they way the drivers behave now when presented with fp24 shaders. After all, according to you, the drivers automaticaly replace all shaders with mixed precision ones.

another quote from u "Drivers can do pure FP24, and mixed mode both......."

the FX cards don't have a fp24 mode, so to run dx9 shaders at full precision they have to use fp32 mode. Please tell me you allready knew this?? And if you did, please explain the quote above.



QUOTE FUNKYMONKEY:

Drumphil.........
HAHAHAHAHAHA you are proved stupid over there also, people overthere dont like your bashing.
I am sorry to say but you are a guy who dont understand simple english.
This is the first time ever i flamed anyone but i dont have any other words. Anyone else can explain him I will give away $ 100 to that person, anyone?
Simple Facts=== Future is Future..
NV 35 Can do FP 24.........True, but slower.
NV 35 can do FP 16/32 combination code ...............true and faster that FP24
NV 35 If needed can do FP32..........True.

If you dont know NV 35 supports Upto 128 bit Floating Precision and RAdeon 9800 pro supports Upto 96 bit precision..
Want any proof of that?
here you go.
http://www.hardocp.com/article.html?art=NDM5

The drivers which are comming out will support Mixmode, I never said that it will do everythink By FP 16/32 Combination only. If Fp24 is forced it will do FP24. If in future FP32 does become reallity it will do FP32, and by then do you think NVIDIA wont comeup with new drivers?
And if anyone will read all the posts you are the one who is not ready t accept the possibilities and facts, anytime if anyone says something about NVIDIA you come bashing.........
NV 35 does has its issues and I am not denying those, you are doubting the capabilities of the cards which are pure technical data that noone can deny and its there in the cards specifications.

If you want direct comparison of these key specifications then here you go

http://www.hardocp.com/image.html?i...EJfMl80X2wuZ2lm
Stop this attitude and try to learn and accept the things as they are...
Noone here cares what you and I say , everyone only cares about the facts.

See and read both the links.....................

No offence to you but please try to accept things and learn....


and


QUOTE FUNKYMONKEY

lets see who is idiot.
I tried to explain you in Laymen's words.
I know that mix mode dosent mean using just FP16. Mixmode will work on any Dx9 game which is backword compatiable with Dx 8.1.
No need to write saperate code for each games with different shaders.
There are total 22 Basic Dx9.0 shaders. And each follows specific algorithm to impliment itself.
Your statement that NVIDIA will need to write a different code for each game is completely wrong.

reever each game comming out will support mix mode.....

If you people dont understand me ask anyone in these forums who knows about this. I know there must be people who know about this. There is no simpler word to explain what i am saying.
Nvidia is not just rewritting shaders for HL2 with Det 5x.xx.
What you can try some DX9 benchmarks itself. Use Det4x.xx in any of DX9 benchmark and then use Det 51.75 you will see the difference.
There is no way NVIDIA is devloping 5x.xx series for only HL2.
16/32 mode will be universal and will work in any game....
If you dont get what i am saying then i am sorry i cannot explain it to ya any better...............
and Drumphil you have said in like 10 threads i will post links to prove yourself but ou havent done it 1 time.
Go on and prove me wrong.

Flaming aside.. is it just me or does he have things very mixed up technically. If i got this wrong im going to get my head checked, cause im sure that thats crazy talk im arguing against. Hate to find out I've lost it..

Here is a link to the thread. http://www.guru3d.com/forum/showthread.php?s=&threadid=65144&perpage=10&pagenumber=7

digitalwanderer
10-02-03, 12:34 PM
Originally posted by Drumphil
and


Flaming aside.. is it just me or does he have things very mixed up technically.
He's mixed up technically on a few points I can see, the first one that springs up is nVidia can NOT do 24FP....it either has to run it in 32FP or drop down to 16FP.

rth
10-02-03, 12:39 PM
he's a 'tard. FX cannot do FP24.

SlyBoots
10-02-03, 01:11 PM
If I'm not mistaken he's also wrong here>

"If you dont know NV 35 supports Upto 128 bit Floating Precision and RAdeon 9800 pro supports Upto 96 bit precision.."

doesn't the 9800 do about 90% of the work at 128bit? Pretty sure I've seen this at B3d, but I'm to lazy to go check
:o

Toaster
10-02-03, 01:21 PM
Originally posted by SlyBoots
If I'm not mistaken he's also wrong here>

"If you dont know NV 35 supports Upto 128 bit Floating Precision and RAdeon 9800 pro supports Upto 96 bit precision.."

doesn't the 9800 do about 90% of the work at 128bit? Pretty sure I've seen this at B3d, but I'm to lazy to go check
:o

I guess he's talking about 128bit color ( 4x FP32 ) and 96bit color ( 4xFP24 )

The radeon does most of the work in FP32 (just like the gfFX) It's the pixelshaders that have a lower precision.

TheTaz
10-02-03, 01:27 PM
Originally posted by Drumphil

~NV 35 Can do FP 24.........True, but slower.
~NV 35 can do FP 16/32 combination code ...............true and faster that FP24
~NV 35 If needed can do FP32..........True.



~Wrong
~Major Wrong... it's slower than FP24... by ALOT... that's the problem.
~Yeah... when DX10 comes out... 1.5 years or more from now...then plus another year for the DX10 games to show up.

Taz

Deathlike2
10-02-03, 02:02 PM
Show me anywhere where I said that fp32 is a standard.. whats that got to do with it... If you run in fp24 mode, the NVIDIA cards DO render it in fp32, because they can't do fp24 (which is one of the BIG reasons they are slower than ATI card when running at dx9 full precision (fp24) ). What exactly is it about fp24 and fp32 that you think I got wrong?? I don't understand your complaint

to quote you: "And yes NV 35 can and will run pure FP32 if any aplication comes out in future........."
.. but don't you say that the drivers will replace all shaders with mixed precision ones??.. Why would that change once fp32 comes in.. How will that be different to they way the drivers behave now when presented with fp24 shaders. After all, according to you, the drivers automaticaly replace all shaders with mixed precision ones.

another quote from u "Drivers can do pure FP24, and mixed mode both......."

the FX cards don't have a fp24 mode, so to run dx9 shaders at full precision they have to use fp32 mode. Please tell me you allready knew this?? And if you did, please explain the quote above.

It is correct that NVidia is being slower in FP32.. that's a given.. the NV3X series really CANNOT run ANY full FP32 program in the future.. that's impossible BY NVidia's current NV3X design

The drivers will most likely convert all shader code to a lower precision (the only true way of gaining performance, with a pretty good chance of losing IQ in the process)

Unless the NV40 ever fixes these flaws, there's nothing to be done about it... the NV3X series is not properly designed for use in DX9 shaders... that's it..

Drumphil.........
HAHAHAHAHAHA you are proved stupid over there also, people overthere dont like your bashing.
I am sorry to say but you are a guy who dont understand simple english.
This is the first time ever i flamed anyone but i dont have any other words. Anyone else can explain him I will give away $ 100 to that person, anyone?
Simple Facts=== Future is Future..
NV 35 Can do FP 24.........True, but slower.
NV 35 can do FP 16/32 combination code ...............true and faster that FP24
NV 35 If needed can do FP32..........True.

If you dont know NV 35 supports Upto 128 bit Floating Precision and RAdeon 9800 pro supports Upto 96 bit precision..
Want any proof of that?
here you go.
http://www.hardocp.com/article.html?art=NDM5

The drivers which are comming out will support Mixmode, I never said that it will do everythink By FP 16/32 Combination only. If Fp24 is forced it will do FP24. If in future FP32 does become reallity it will do FP32, and by then do you think NVIDIA wont comeup with new drivers?
And if anyone will read all the posts you are the one who is not ready t accept the possibilities and facts, anytime if anyone says something about NVIDIA you come bashing.........
NV 35 does has its issues and I am not denying those, you are doubting the capabilities of the cards which are pure technical data that noone can deny and its there in the cards specifications.

If you want direct comparison of these key specifications then here you go

http://www.hardocp.com/image.html?i...EJfMl80X2wuZ2lm
Stop this attitude and try to learn and accept the things as they are...
Noone here cares what you and I say , everyone only cares about the facts.

See and read both the links.....................

No offence to you but please try to accept things and learn....

NV35 can't do FP24.. only FP16 and FP32.. if FP24 is used.. FP32 will be used unless the NVidia compiler forces FP16..

The combination code MAY be faster.. but not proven yet

NVidia does have 128-bit precision and ATI does have 96-bit precision.. but ONLY for their PS hardware.. the VS hardware is all 128-bit...

However, NVidia won't have good performance with high precision.. so it doesn't matter if you have MORE precision, it's not going to run WELL.

MORE != FAST

More precision incurrs more math being done.. which is inherently slower (using higher precision generally increases the work done, thus requiring more power)

Unless the NVidia hardware was slightly faster.. it could offset the FP32 vs FP24 performance penalties.. however the lack of registers (and lack of other things, which you should check out Beyond3D to learn more of) prevents the NV3X Pixel Shaders from running at full efficiency.

Part of the NV3X hardware design is to share the shaders with the texture pipelines (I think)... basically, there is a a "holdback" on performances because of the fact that its sharing... if the hardware itself was MUCH faster, that may offset the sharing penalties... however, that is not so. It would've been smarter to have a dedicated shader pipeline.. as this should improve performance drastically...

You can't fix hardware with software in general... you can only do so much.

lets see who is idiot.
I tried to explain you in Laymen's words.
I know that mix mode dosent mean using just FP16. Mixmode will work on any Dx9 game which is backword compatiable with Dx 8.1.
No need to write saperate code for each games with different shaders.
There are total 22 Basic Dx9.0 shaders. And each follows specific algorithm to impliment itself.
Your statement that NVIDIA will need to write a different code for each game is completely wrong.

reever each game comming out will support mix mode.....

If you people dont understand me ask anyone in these forums who knows about this. I know there must be people who know about this. There is no simpler word to explain what i am saying.
Nvidia is not just rewritting shaders for HL2 with Det 5x.xx.
What you can try some DX9 benchmarks itself. Use Det4x.xx in any of DX9 benchmark and then use Det 51.75 you will see the difference.
There is no way NVIDIA is devloping 5x.xx series for only HL2.
16/32 mode will be universal and will work in any game....
If you dont get what i am saying then i am sorry i cannot explain it to ya any better...............
and Drumphil you have said in like 10 threads i will post links to prove yourself but ou havent done it 1 time.
Go on and prove me wrong.

Um.. mixed mode is only applicable to HL2's engine... when did this "new mode" appear in all DX9 games... maybe its the fact that Tome Raider (a DX9 game) has it's own special mixed mode "which is the CG compiler mode" which renders the game with decent performance... however the DX9 mode makes the game run with "missing visual effects and lower performance"...

The "mixed mode" thing has appeared in quite a lot of things.. including Doom 3... special paths must be written for NV3X cards because there using FP32 for everything will make it run at unacceptable levels.

NVidia may be rewriting shaders in ALL games that use PS 2.0... NVidia will develop "optimizations".. not JUST HL2 (I believe some global optimizations will be made, as well as game specific optimizations)

FP16/32 is not "universal".. I don't see how, FP16 is the minimum spec for "low precision" and FP24 is the minimum spec for "default precision"... More developers are developing with FP24 in mind.. and all kinds of precisions in NVidia's hardware (in NVidia's case, to produce acceptable performance)

Mixed mode is something that NEEDS to be CREATED for the NV3X series.. it is something that should NOT have been created in the first place. MS let a precision extension for lower precision for NVidia's sake, as it DOES NOT affect ATI cards (ATI cards run ONLY at FP24, so you'll get the same precision on all accounts)...

Developers don't really want to do this.. but they technically HAVE to support this mode.. because it runs unacceptably on NV3X hardware and there are many NVidia users out there... they can only do so much to make everyone happy. Developers are really under NO obligation to have to done this in the first place.

If hardware was naturally slow.. it's not worth to code it unless there was some compelling reason (like HL2 being the most likely game of the year)..

If hardware was naturally fast... and developers deliberately kill its performance.. that's another story...

Example of games affected by the first case:

HL2
Tome Raider:AOD

Examples of games affected by the second case:

Neverwinter Nights (it may not have been deliberate, but they killed performance on the ATI cards... at least Bioware plans to work on fixing that)

Let me sum it up simply for you. If a developer uses pure FP24 code (with DX9 PS2 specs), NVidia's hardware will suffer in trying to run it in FP32 (as they are forced to by default).. Unless NVidia forces ALL FP24 operations to be converted to FP16 (or less), the code will NOT run at the same speed as the ATI hardware...

If forcing all operations to be FP16 (or less) has equivalent performance to ATI's FP24, you are SURE to see some IQ differences (and most likely you will see them)

Using FP32 for ALL operations is technically UNUSABLE if you use a LOT of it... you can use it ONLY if you are doing VERY LITTLE of it.

Deathlike2
10-02-03, 02:34 PM
I'll do the smart thing and just link it..

http://www.beyond3d.com/reviews/albatron/gffx5900pv/index.php?p=4
http://www.beyond3d.com/previews/nvidia/nv35/index.php?p=8
http://www.3dcenter.org/artikel/cinefx/index_e.php -The best analysis

If the 3dcenter one doesn't get you.. nothing will.. lol

StealthHawk
10-02-03, 05:38 PM
He doesn't have a clue what he's talking about.

NV3x does not support FP24 whatsoever. There are three precisions that NV3x supports and can use: FX12(not legal in DX9 shaders), FP16(partial precision), and FP32(default precision).

For NV35 FX12 seems to be basically the same speed as FP16.

Drivers cannot "enable" mixed mode. Mixed mode is achieved through code paths included in the game or with NVIDIA software engineers hand coding replacement shaders. This is not something that a compiler dynamically handles.

NV35 will be able to technically use FP32 in future games, but it will be utterly unplayable and so this isn't a viable choice whatsoever.

ChrisRay
10-02-03, 07:33 PM
Man. That is some broken english there.. And I thought my own paragraphs were hard to read...

hithere
10-02-03, 08:17 PM
I think the basic difference between the two cards in this department, performance-wise, is that the R300 series has an 8-pipeline, all the time, architecture....meaning that each of its pixel pipes has one shader with a precision of 24-bit available to it.

The FX series, on the other hand, acts like an 8-pipe card when handling partial precision shaders (16-bit), but acts more like a 4-pipe card when running full precision (32-bit)...it sort of dynamically allocates regiseter usage depending on the amount of precision the shader requires. Bottom line is that there are more registers available for shaders on the R300 series no matter how you look at it, and the only way the FX can keep up in performance is to run a lower precision shader (16-bit or lower) vs the R300's 24-bit. The clock speed of the FX is the only thing, in my mind, that keeps it from being clobbered even worse.

stevem
10-02-03, 11:02 PM
Originally posted by hithere
The FX series, on the other hand, acts like an 8-pipe card when handling partial precision shaders (16-bit), but acts more like a 4-pipe card when running full precision (32-bit)...it sort of dynamically allocates regiseter usage depending on the amount of precision the shader requires.
That's an interesting interpretation of NV30/5 pipeline function. It's a 4x2 except for non-color writes. Both NV30/5 & R3x0 operate on quads. No pipeline conformation change, as such, occurs between full & pp_hint (fp) modes on NV30/5. Register limitations are magnified, however...

Originally posted by StealthHawk
There are three precisions that NV3x supports and can use: FX12(not legal in DX9 shaders)
Nvidia's new shader optimising driver core allows for insertion of int code for DX9 shaders as it's done at the assembly level... Neatly subverting DX9 shader spec, don't you think...?:)

Edit:typos

StealthHawk
10-03-03, 05:24 AM
Originally posted by stevem
Nvidia's new shader optimising driver core allows for insertion of int code for DX9 shaders as it's done at the assembly level... Neatly subverting DX9 shader spec, don't you think...?:)

Edit:typos

I have no idea what Det50's shader compiler is doing and how much of it is automatic. All I know is that from what we've seen with 51.75 I don't have my hopes up in terms of performance or IQ.

hithere
10-03-03, 10:25 AM
Originally posted by stevem
That's an interesting interpretation of NV30/5 pipeline function. It's a 4x2 except for non-color writes. Both NV30/5 & R3x0 operate on quads. No pipeline conformation change, as such, occurs between full & pp_hint (fp) modes on NV30/5. Register limitations are magnified, however...


You're right, I shouldn't have used the number of pipes...but the end result, in my under-educated analysis, is that you magnify the register usage when you go to 32-bit FP, and the card becomes slower at those operations (by a factor of 2?). There's no magical change in its pipeline architecture, but the end result is that you're waiting on the shader engine to do the work anyway.