ATi response to Doom 3 Benchmarks

[grecco_julio]

TeamRadeon.com:

"TI Response to Doom 3 Benchmarks
Published: 7/23/2004
Author: Team Radeon


Hear what ATI have to say about the latest Doom 3 benchmarks.


One of our staff contacted ATI to inquiry about the highly debated Doom 3 benchmarks that were recently released, here is what was asked:

"I'm just emailing you because I have alot of confused gamers from both Team Radeon and x-3Dfx about what's really happening in Doom3. What they've been presented paints a relatively poor picture on the x800 Pro and x800 XT PE prompting many fellow gamers (about 400 on Team Radeon and 300 or so on x-3Dfx) to question wether or not they should cancel there pre-orders."

ATI response:

"Hi Richard - this is a non issue - Doom 3 isn't even available yet, and we all know that some of our competitors use partial precision where possible. We expect to have updated drivers available in the coming weeks."

Chris Hook

Check out our forums to join the current discussion about the Doom 3 benchmarks."

[IbaneZ]

I thought he'd say something like:

"Well, sorry, Doom 3 will always run like crap on our hardware. So buy the competitors cards if you're going to play that game.
And no, we have no plans on updating our Open GL drivers either.
Good day."

A lot of damage control in progress now, so i guess we'll just have to wait and see what happens. 8)

I sure hope they'll bring us some magic and make the life easier for my X800 Pro. :)

[hovz]

actually its a very real issue cuz a lot of people will buy nvidia cards because of doom 3.

[grecco_julio]

I trust them. They wouldn't be dumb aswell to don't improve their drivers to Doom 3, they'd lose sellings for NVIDIA.

[digitalwanderer]

It's not a matter of trust for me as I know ATi will come thru, the big question for me is WHEN? :?

[grecco_julio]

"Hi Richard - this is a non issue - Doom 3 isn't even available yet, and we all know that some of our competitors use partial precision where possible. We expect to have updated drivers available in the coming weeks."

right after the release of Doom 3 :twisted:

[digitalwanderer]

"Hi Richard - this is a non issue - Doom 3 isn't even available yet, and we all know that some of our competitors use partial precision where possible. We expect to have updated drivers available in the coming weeks."

right after the release of Doom 3 :twisted:
PROBLEM: Doom3 isn't coming out "in the coming weeks", it's pretty much coming out NEXT week! :?

I have a gut feeling that the openGL re-write will NOT be ready when D3 hits the store shelves, my big question mark is how long between D3 hitting the shelves until the openGL re-write is released.

THAT will be a huge determinent for a lot of people.

[hovz]

its not a matter of ati not trying, its can ati rly bridge the gap with just drivers? and dont u expect that nvidia will also improve the performance? theres rly not much ati can do. they will most likely always be behind in doom 3

[grecco_julio]

SInce I'll be buying'n buying a X800 XT-PE and playing Doom 3 in the 1 week of September I'm not worried about it.

[digitalwanderer]

its not a matter of ati not trying, its can ati rly bridge the gap with just drivers? and dont u expect that nvidia will also improve the performance? theres rly not much ati can do. they will most likely always be behind in doom 3
ATi can bridge a lot of that gap with a new openGL ICD, IMHO....but we won't know until we know. :)

[alanity]

we all know that some of our competitors use partial precision where possible.That really makes me cringe, wouldn't it be great if these companies relied on their own strengths instead of milking any bad rep their competitors get for all it's worth? 'Cause, like, it would be really cool to try and catch one of the pigs that would be flying around and ride it.

[PowerK]

Typical PR statement...

I personally think the results from that DOOM3 benchmark is pretty much maxed out performance of Radeon graphic cards. ATI had time to improve their OpenGL driver before they went to id office for official benchmarks, and before DOOM3 releases in August.

Whatever driver ATI used for the benchmark, I think it already provided considerable improvement over Catalysts that are already released in public, and we have. ( *If* this was not the case, then... it's pretty much obvious that ATI's OpenGL driver team is simply hopeless. They've had more than enough time to improve OpenGL driver, and they *knew* DOOM3 was coming ).

IMHO, very minor/slight improvement will be provided by near future drivers, but nothing to write home about.

[GrapeApe]

THAT will be a huge determinent for a lot of people.

Well even more of a determinent IMO would be the availability of cards by that time. Reallly, there's such a lack from both sides I think the actual release of D3 won't have as big an impact as getting the OGL 'fix' ready and benched/reviewed before there are actually 'reasonably priced' cards available for purchase.

I plan on playing it alot on a P4 laptop with an mR9700 while on vacation, so nothing would change that purchase plan (buying next week depending on IBM/DELL salespeople's responses), I'm sure a few others are in a similar boat, it's all fine and good to talk about it being a killer app, but if price and availability don't improve the MAJORITY of buyers will play D3 the first go through on their current hardware, and then enjoy it once again with the settings cranked once they replace their card.

Anywhooo, just a thought.

[digitalwanderer]

It's a good thought, that will be a factor...but I think that the availability of cards on both sides is actually pretty decent (as long as you're not too picky on the flavor ;) ) and it should improve dramatically right about the time D3 releases.

I think everyone knows in the hardware business that it is now officially x-mas in July and it's a "you snooze you lose!" situation about having products on the shelf. (I'm really started to get excited about the whole "rush to upgrade for D3" thing, it's gonna bring a WHOLE lot of newbies into the community! :D )

[GrapeApe]

I think everyone knows in the hardware business that it is now officially x-mas in July and it's a "you snooze you lose!" situation about having products on the shelf.

True, I just wish it pushed people to release PCIe and Socket 939 solutions, not to release stuff that should've been here a while back.

Oh well I'll be laptop bound for the month of Aug anyways, and playing D3 should be interesting to see how it handles things. The X600 based solutions aren't out there from what I've seen and purchase date is near the end.

The Flight East should be interesting/entertaining enough, if only for the shocked expressions of the other passangers at the evil/disgusting images oeminating from the screen; and maybe when I'm done with all the pr0n I'll play a little D3 too and really mess with their mellons. :twisted:

[jvd]

Heh. We will have a new driver for doom3 It wont make up the 10fps . But it will make up enough so the x800pro doesn't look that bad anymore

[logen999]

an update from chris Hook.....


...And btw, let's not lose sight of the fact that ATI performance isn't
relatively poor at all. I think Kyle himself said that even the X800pro
delivered 'great' performance, and Carmack said in the HardOCP article that
there's more to consider than just frame rate. The frame rate difference
even today is so minor, it's impossible to tell without diagnostic tools -
ie: the end user experience isn't affected. And with ATI you get
full-precision enabled all the time - we don't do PP (on R3XX and R4XX) like
some of our competitors.

It's also important to note that most of today's games play faster on ATI
hardware, and you can expect that to extend to other 'big title' games
expected this summer.

Chris

Chris Hook

thanks to my pall ElMoIsEviL from TR. hopefully Carmack will respond too. i'd liek to hear what he has to say. :wink:

[thop]

we all know HL2 is gonna be faster with ATI cards, but who cares about HL2 at the moment when everyone's talking about D3.

[Razor1]

THAT will be a huge determinent for a lot of people.

Well even more of a determinent IMO would be the availability of cards by that time. Reallly, there's such a lack from both sides I think the actual release of D3 won't have as big an impact as getting the OGL 'fix' ready and benched/reviewed before there are actually 'reasonably priced' cards available for purchase.

I plan on playing it alot on a P4 laptop with an mR9700 while on vacation, so nothing would change that purchase plan (buying next week depending on IBM/DELL salespeople's responses), I'm sure a few others are in a similar boat, it's all fine and good to talk about it being a killer app, but if price and availability don't improve the MAJORITY of buyers will play D3 the first go through on their current hardware, and then enjoy it once again with the settings cranked once they replace their card.

Anywhooo, just a thought.

Nvidia cards are available, you can get ultras from the PNY, EVGA, gts are in stores. The only card that you can't get is the x800xt pe. Actually if you take alook around about 2 days after the doom 3 benches were released alot of estores ran out of 6800 gts. This benchmark had a huge influeance.

Nice laptop btw mr9700 is a really good card.

I don't think the new drivers that will equal the nvidia cards performanec will be out anytime soon for ATi, as someone said earlier. If they didn't do them for the Doom 3 benchmarks, they must have been doing something else. They knew the impact of Doom 3 and they knew thier low scoring in ogl games.

[logen999]

we all know HL2 is gonna be faster with ATI cards, but who cares about HL2 at the moment when everyone's talking about D3.

I wouldn’t be so quick to draw conclusions. At this point in the game Nvidia has really showed off its shader muscle. I expect (hope) both cards will handle the game equally well since it’s a much more unified Dx 9 code path.

QUIT STALKING ME LOGEN!!!!

Sorry, don't mind me.... :oops:

:lol: :twisted:

[Scali]

I think ATi could improve performance considerably... all they have to do is cheat. Carmack pretty much said between the lines that NVIDIA is doing shader replacement. There's nothing stopping ATi from doing the same... And perhaps that is why ATi is not really worried, and are confident that the driver update will 'fix' the performance.

[Quitch]

I wonder if the Doom 3 benchmarks represent "better at OpenGL" or "designed for Doom 3".

[Trawler]

Could be a while until the full OpenGL ICD is avilable... perhaps ATI will release a MiniGL D3D wrapper for the interim? :twisted:

[peh]

"The frame rate difference even today is so minor, it's impossible to tell without diagnostic tools - ie: the end user experience isn't affected."

-- Chris Hook

I'm not sure that you'd need fraps to determine the difference between 36 fps on a 6800GT and 21 on a X800pro.

Additionally the ~17 fps margin that the Ultra has over the XTPE @ 16x12 8xAF would probably be naked-eye detectable.

Of course these are just 2 of several tested scenarios

(please don't flame me about how quickly the human eye can perveive changes ... im just stating my opinion)

His implication the the ATI OGL rewrite (or at least heavy D3 optimisation) would be finished soon seems dubious given how many weeks ahead the driver version they were using seems to be, and how many years the OGL rework itself is taking.

His points about the XTPE winning for the majority of other games is obviously correct, but doesnt really matter when people are changing their card purchase decisions on the basis of this game alone.

the _pp comment was hilarious as well (and not because of its inaccuracy either)

and no, im not sure why i just checked a PR statement for accuracy...

--peh--

[AlStrong]

how many people are actually rewriting the opengl drivers at ati :?:

[mono]

His implication the the ATI OGL rewrite (or at least heavy D3 optimisation) would be finished soon seems dubious given how many weeks ahead the driver version they were using seems to be, and how many years the OGL rework itself is taking.

Quoted for emphasis. ATI has been lagging in OpenGL for years, not without outcry. I wouldn't expect to see some miraculous short-term OpenGL rewite.

[Ostsol]

His implication the the ATI OGL rewrite (or at least heavy D3 optimisation) would be finished soon seems dubious given how many weeks ahead the driver version they were using seems to be, and how many years the OGL rework itself is taking.

Quoted for emphasis. ATI has been lagging in OpenGL for years, not without outcry. I wouldn't expect to see some miraculous short-term OpenGL rewite.
Wasn't it a general lack of stability and compatability that was the cause of most of that outcry?

[micron]

Wasn't it a general lack of stability and compatability that was the cause of most of that outcry?
Oh yea....I dont think IQ or speed was ever an issue.

[trinibwoy]

Wow I'm impressed by the faith in ATI on this board. I'm all for a performance increase in OGL on R420 but damn this thing is getting religious!!! Did Dave Orton form a cult? :lol:

[Ostsol]

Wasn't it a general lack of stability and compatability that was the cause of most of that outcry?
Oh yea....I dont think IQ or speed was ever an issue.
Well, IQ is up to the hardware (unless the driver forces it to cut corners or causes artifacts -- though the latter is stability/compatability issue). Speed kinda has been an issue, though. My point was that speed probably wasn't formost on peoples' minds back when stability was an issue.

[bloodbob]

The X800 Pro might get a massive preformance increase but its doubtful if any card will. Hopefully the next engine will use GLSL and if so I hope the ATI cards should run that rather well by then. As we often here about those X-Dec guys.

Still I need a new processor for doom anyway :/

[DemoCoder]

D3 does seem rather CPU limited. I'm guess the x-box version has a lower amount of shadow trees.

[DSN2K]

Wow I'm impressed by the faith in ATI on this board. I'm all for a performance increase in OGL on R420 but damn this thing is getting religious!!! Did Dave Orton form a cult? :lol:

just think over the last few years expectation's have risen for ATI and people expect them to sort it out. :wink:

[solobird]

1/ JC has removed the NV30 Path, all NV30 and R300 or above are running at ARB2, so there is no precision issue anymore

2/ JC stated there is no trick in their official benchmark

3/ JC also adressed that if color mipmap is used in the test, R3xx and above will drop even more

4/ I am curious why he did not mention anything about 6800 vs. X800

So it is quite obvious this statement is useless but for marketing.

As some friend said, "if JC think FP16 is quite enough for the game, and there is no IQ issue, then nobody should complain about this, since it the the JC's choice. And the precision adoption is only JC's option, not us".
:roll:

[jvd]

Solobird have you read anything posted on this forum ? Because if you did u would kknow


1 ) in email Carmack said the nv3x path was mostly roled into arb2 and nvidia moved the optimizations into the driver. Thus they didn't need the nv3x path anymore as nvidia was already through the driver forcing the lower percision.

2) JC stated that if the testwas to change nvidia's drivers would fall off the fast track. THat means if anything changed like the shaders they drivers would no longer hav the optimizations on and would loose speed.

3)Right he said if you knew what you were looking for like him it could be a problem.

4) Why would he . ITs a 10fps lead .

I agree his statements are only good for nvidia marketing .

Till you get to him saying framerates aren't the only thing you should think about. Things like dual monlex connectors , two slot cooling and other things weighing in against diffrent cards .

As I say if Jc said for us we can't see a diffrence with ati's filtering and he isn't complaing then either should we .

[solobird]

jvd, thanks for your reminding, however

1/Just as I said above, the precision which is adopted by JC is not a question for us to talk about. Since it is the decision by JC. If he thought FP16 is good enought for us, then it is OK!

2/I think the "Fall off the fast path" is like the scene, which happened in 3DMark 2003 patch. It is a driver level optimization and a swordplay.

3/Since there is not an apple-to-apple competition anymore, ATi does the adaptive texture analysing & filtering and NV makes the mixmode precision optimization, it will hurt each of them if the color mipmap is enabled and the code is changed. It seems any of us should not complain about these tricks anymore, right? :P

4/ I think 10fps is quite huge, since the fastest card can only run the game at 73fps at 1024*768 4XAA 4XFSAA


On all accounts, can we call it a fair match?
:shock:

[jvd]

1/Just as I said above, the precision which is adopted by JC is not a question for us to talk about. Since it is the decision by JC. If he thought FP16 is good enought for us, then it is OK! we don't know if its done by jc . He said if things were changed nvidia would fall off the fast track. So one of the things they may be doing is replacing shaders with less complex ones and with fp 16 or interger instead of what carmack wants. We don't know though . Perhaps we never will know

/I think the "Fall off the fast path" is like the scene, which happened in 3DMark 2003 patch. It is a driver level optimization and a swordplay.
No I'm pretty sure (And many others here) that it has to do with optimizations. This may be an issue where nvidia is faster in standard doom 3 because of the optimizations but will fall behind in mods because the shaders havnen't been optimized.




3/Since there is not an apple-to-apple competition anymore, ATi does the adaptive texture analysing & filtering and NV makes the mixmode precision optimization, it will hurt each of them if the color mipmap is enabled and the code is changed. It seems any of us should not complain about these tricks anymore, right?

Well talk to Chrisray he doesn't believe so.

Personaly I don't care. I hav ea 6800gt clocked to ultra speeds in the house and a x800xt pe . So far the only reason to take the 6800 out of my fathers pc is doom 3.



4/ I think 10fps is quite huge, since the fastest card can only run the game at 73fps at 1024*768 4XAA 4XFSAA

There are other factors . I can get 4x fsaa image quality from ati while only running at 2x fsaa performance.

On all accounts, can we call it a fair match?
I personaly don't care . I believe in future drivers ati will make back some of that performance. I think nvidia will allways have the lead but not by as much as it currently does .

We don't know what those drivers really are from ati. We don't know what was really going on with nvidia's drivers.

We don't know how these cards will perform on other time tests or in the actual game and we only have benchmarks from one source.

So i'm going to wait and see what happens when other sites do reviews.

[solobird]

urh, it will make sense to wait for other reviews. :P

[Smurfie]

I don't understand why people will kick a big fuss over PP. In every day programming, we save bits where possible, and if sufficient precision is achieved using fewer bits, do it.

How many people still use 32bit integer for boolean variables? Yes, 32 bit is higher precision, but hell, that's many bits for 1 and 0.

Likewise, if JC thinks he has enough bits for his engine, I think nobody has any grounds for criticism. But should PP prove to produce precision errors and cause visual effects, yes, we can call JC out on this.

[radar1200gs]

I'd have to dig up the actual carmack quote (it is somewhere here on beyond3d though), but, OPenGL does allow precision hints and carmack wrote his shaders with those precision hints in place in a lot of (not all) cases.

So, I fail to see how ATi PR can claim it is their competitor running partial precsion when (by the implication of the PR rep) they should not be - the game was designed by the programmer to take advantage of part of the OpenGL specification!

ATi should explain why their products don't fully support the specification (no partial precsion) before criticising other IHV's for correctly using it in a game designed for its use!

[kkevin666]

What people have wanted from ATI drivers over the last 2 years

STEREO SUPPORT
PER GAME AA/AF SETTINGS without needing to open control panel everytime
BETTER OPENGL DRIVERS

What people have got over the last 2 years

-----------------------
See where im going with this?

Do not realistically expect any opengl rewrite anytime in the future- ATI management dont seem to realise this is a major factor in people not buying their cards .

[PatrickL]

You are confused: your needs are not People needs :)

[Mariner]

ATi should explain why their products don't fully support the specification (no partial precsion) before criticising other IHV's for correctly using it in a game designed for its use!

Preposterous! Criticising ATI for not having (unnecessary) Partial Precision is just ridiculous.

I know there is plenty of ATI flag-waving going on here, but getting your green flag out as well benefits nobody.

The PR guff from ATI is equally ridiculous. Claiming, "Oh, we have higher precision" is just echoing the PR stuff coming out of NV the other year when trying to defend their lacklustre NV3X series. It just adds fuel to the fanboi fires.

We know the NV chips are practically designed around Doom3 so it shouldn't come as any surprise that they outperform their ATI counterparts. Carmack says "Scores will probably improve somewhat with future driver releases" and I'd expect the ATI cards to get closer to NV in performance terms as we know the current ATI OpenGL drivers leave a lot to be desired. I doubt NV will be able to eke much more performance out of their drivers which are widely regarded as very good already. I still expect NV to have a reasonable lead, however.

Ultimately, regardless of the length of these fanboi pissing contests both NV4X and R4XX cards should provide very good performance in Doom3. I'd say the performance of the forthcoming mainstream (8-pipe) cards is more important than the high-end monsters. After all, these will sell in massively greater numbers.

[radar1200gs]

Hey Mariner, don't blame me!

ATi are the ones insinuating nVidia is gaining an unfair advantage from partial precision.

Partial precision is part of both the OpenGL and Direct3D specifications.

Instead of of whining about products that can use partial precision actually using it when the programmers allows it to be used ATI needs to justify why partial precision is not necessary for them (if its not needed why complain???)

Either that or shut their cakeholes...

[whql]

I didn't see them complaining, I saw them reminding people that they run in high precision all the time.

[radar1200gs]

I didn't see them complaining, I saw them reminding people that they run in high precision all the time.
Bullshit!
ATI response:

"Hi Richard - this is a non issue - Doom 3 isn't even available yet, and we all know that some of our competitors use partial precision where possible. We expect to have updated drivers available in the coming weeks."

Chris Hook

[whql]

And they are highlighting that they don't - if you want high precision rendering the only guarantee you have of getting it is with ati. We've all seen the far cry stuff that has everything coded in partial precision, so its a fair guess that this is used elsewhere and given carmacks comments before it seems that he has as well - so if you want full precision (which is one of nvidia's big marketing points) you can only be assured of getting that with ati.

[Farid]

D3 does seem rather CPU limited. I'm guess the x-box version has a lower amount of shadow trees.

Yes, indeed (http://img.kult-mag.com/photos/00/00/44/17/ME0000441752_2.jpg). The Light/Shadow system has been scaled down for the Xbox.

[radar1200gs]

What ATi feels about precision is irrelevant. What JC does within OpenGL specs in his game is relevant.

The quote is from me. Nvidia probably IS "cheating" to some degree, recognizing the Doom shaders and substituting optimized ones, because I have found that making some innocuous changes causes the performance to drop all the way back down to the levels it used to run at. I do set the precision hint to allow them to use 16 bit floating point for everything, which gets them back in the ballpark of the R300 cards, but generally still a bit lower.

Removing a back end driver path is valuable to me, so I don't complain
about the optimization. Keeping Nvidia focused on the ARB standard paths
instead of their vendor specific paths is a Good Thing.

The bottom line is that the ATI R300 class systems will generally run a
typical random fragment program that you would write faster than early NV30 class systems, although it is very easy to run into the implementation
limits on the R300 when you are experimenting. Later NV30 class cards are
faster (I have not done head to head comparisons with non-Doom code), and the NV40 runs everything really fast.

Feel free to post these comments.

John Carmack

[whql]

What ATi feels about precision is irrelevant.
So you agree that nvidia flogging 32 precision for all it worth is irrelevant? all ati are saying is if you want to be assured of high precision you can only get that 100% of the time with ati.

[radar1200gs]

FP32 is available and can be marketed and used. It is a higher precision than ATi is capable of. Like any feature it can (and will be) marketed.

That does not exclude usage of FP16 (unless nVidia say this shader uses FP32 and it turns out to use FP16).

The reality is FP16 will prove sufficient at least 90% of the time for a good while to come. If it doesn't prove sufficient FP32 is there to be used.

[silence]

lol. i dont usually comment, but radar really makes me laugh in this topic....
:D

[DeanoC]

ATi should explain why their products don't fully support the specification (no partial precsion) before criticising other IHV's for correctly using it in a game designed for its use!

ATI support both the D3D and OpenGL partial precision specifications to the letter.

The specification says that when used the calculations must be a MINIMUM of FP16, which ATI pass 100%.

Its even enshrined in the OGL original specifications (which is 13 years old), that no operations have to be bit accurate as long as they meet the minimum standards set in the specification. It has to be a minimum specification else it would require every new hardware revision to support a legacy pipeline for 8bit fixed, 12 bit fixed, FP16, FP24, FP32 and a few others.

[radar1200gs]

lol. i dont usually comment, but radar really makes me laugh in this topic....
:D
How so? Have I misquoted anyone? said anything untrue?

[Scali]

D3 does seem rather CPU limited. I'm guess the x-box version has a lower amount of shadow trees.

Makes me wonder if Doom3 is actually doing too much work on the CPU (I believe skinning and shadowvolume generation)... 3dmark03 was criticised for doing everything on the GPU, but perhaps that was actually the better solution?
Well, I guess as usual, the truth will be somewhere in between.

[radar1200gs]

ATi should explain why their products don't fully support the specification (no partial precsion) before criticising other IHV's for correctly using it in a game designed for its use!

ATI support both the D3D and OpenGL partial precision specifications to the letter.

The specification says that when used the calculations must be a MINIMUM of FP16, which ATI pass 100%.

Its even enshrined in the OGL original specifications (which is 13 years old), that no operations have to be bit accurate as long as they meet the minimum standards set in the specification. It has to be a minimum specification else it would require every new hardware revision to support a legacy pipeline for 8bit fixed, 12 bit fixed, FP16, FP24, FP32 and a few others.

While I don't know for certain, I'm fairly confident that ATi does exactly what they do under D3D for OpenGL partial precision - ignore the hints altogether. Perhaps one of the ATi guys can elaborate.

Obviously you can do "partial precision" using full precision, but that won't give you any speed benefits. nVidia's way of doing partial precsion does give speed benefits. No use whinging about a feature because you (1) don't support it, (2) support it in a half-assed fashion (3) your competition supports it better than you. That doesn't make it any less a part of the spec or any less able to be used.

[Martin Eddy]

The only thing is, it wasn't part of the spec until NVIDIA pushed microsoft and the ARB to add it, because there performance sucked in full precision.

[silence]

lol. i dont usually comment, but radar really makes me laugh in this topic....
:D
How so? Have I misquoted anyone? said anything untrue?

well, if i wanted to read NV PR i would go to nvidia.com.
i simply find it funny for anyone to fight for so long over this kind of things.
i had big fight with some guys over one croatian gaming forums over something like this...i really dont see what you get in defending/attacking any IHV like this...unless you are payed by one?

thats all.
:D

[Unit01]

lol. i dont usually comment, but radar really makes me laugh in this topic....
:D
How so? Have I misquoted anyone? said anything untrue?

Oh dear you don't see it? :?

ATi should explain why their products don't fully support the specification (no partial precsion) before criticising other IHV's for correctly using it in a game designed for its use!

What?

If i say that you're a human being am i whining that you are that or just pointing out the fact that you are?

For a company that disses FP24 24/7 and then mostly uses FP16 don't you find it ironic or funny at all? <-Information doesn't coincide with bias.....bzzzz discard...proceed

D3 does seem rather CPU limited. I'm guess the x-box version has a lower amount of shadow trees.

Makes me wonder if Doom3 is actually doing too much work on the CPU (I believe skinning and shadowvolume generation)... 3dmark03 was criticised for doing everything on the GPU, but perhaps that was actually the better solution?
Well, I guess as usual, the truth will be somewhere in between.
Isn't the 3dmark03 test pretty static?
I thought that CPU impact on 3dmark03 was very very little. No sound no physics no nothing almost :/
Or am i wrong? Can someone clue me in here?

[DeanoC]

While I don't know for certain, I'm fairly confident that ATi does exactly what they do under D3D for OpenGL partial precision - ignore the hints altogether. Perhaps one of the ATi guys can elaborate.

ATI only support one precision throughout, so its obvious that any partial precision is supported at the same precision as any variable.


Obviously you can do "partial precision" using full precision, but that won't give you any speed benefits. nVidia's way of doing partial precsion does give speed benefits. No use whinging about a feature because you (1) don't support it, (2) support it in a half-assed fashion (3) your competition supports it better than you. That doesn't make it any less a part of the spec or any less able to be used.

No partial precision is a decision ATI believe saves them enough silicon that they can speed up the single precision pipeline. What NVIDIA fans choose to ignore that multiple precisions is only faster IF the cost of supporting more than one is less than the gain you could get by sticking all those gates optimising a single path.

Nobody has shown that supporting multiple precision is actually a winner at the chip level. ATI haven't needed partial precision for the R300 and R400 generations and that won't be changing in future. Indeed if NVIDIA don't drop the whole multiple precision idea, they may start to suffer compared to what single precision designs give you.

Maybe you should look at highspeed designs in other areas of high performance computing, you may be surprised how few support multiple precision. Its not really even a CPU feature anymore and often they 'emulate' it for compatibility with the CPU ISA at the hardware level, almost all modern CPUs only support 1 FPU precision, 1 integer precision and 1 vector ALU precision. They support pack/unpack units that make it appear they support FP32/FP64/FP80, I16/I32/I64 etc.

[Xmas]

Maybe you should look at highspeed designs in other areas of high performance computing, you may be surprised how few support multiple precision. Its not really even a CPU feature anymore and often they 'emulate' it for compatibility with the CPU ISA at the hardware level, almost all modern CPUs only support 1 FPU precision, 1 integer precision and 1 vector ALU precision. They support pack/unpack units that make it appear they support FP32/FP64/FP80, I16/I32/I64 etc.
I agree that single precision is the way to go, except maybe a few special functions like rsq or sin.
NV40 basically is a single precision design, but with the pack/unpack units placed between the register file and the ALUs, so registers can be packed, too.

[solobird]

DeanoC

I do not totally agree with some of your words. I admit that single precision can save silicons, but there is one thing, truely important, that the memory bandwidth and the total amount of memory are limited, therefore it is quite obvious for a programmer to think about the memory utilization. As us sofeware programmer we always try to optimize our code to consume as less memory as we can and to make it as faster as we can.

In game programming, multiple precision will save valuable bandwidth and make it for other features or functions and it provides more flexible way for the programmer to optimize their code. You know what, if the precision is fixed on one, then, some years later, more and more programmer will lose their sense in these optimization skills.

[jpaana]

OPenGL does allow precision hints and carmack wrote his shaders with those precision hints in place in a lot of (not all) cases.


ARB_fragment_program allows only all reduced precision for a single shader with "OPTION ARB_precision_hint_fastest;" or without it full 24+ bit precision, so you don't have control if you want a single operation to be high precision.

NVidia extension to ARB_fragment_program, namely NV_fragment_program_option, allows partial precision temporaries ("short temp" instead of "temp" for full precision) and added R, H and X suffixes to instructions like older NV_fragment_program, but is again NVidia proprietary (although ATI could just ignore the hints).

[radar1200gs]

Yes, I know that. Whats your point?

Partial precision: ARB_precision_hint_fastest
Full precision: ARB_precision_hint_nicest

These work on entire shaders.

[jpaana]

Yes, I know that. Whats your point?

Partial precision: ARB_precision_hint_fastest
Full precision: ARB_precision_hint_nicest

These work on entire shaders.

As I said above, "you don't have control if you want a single operation to be high precision." which might be the case if you want to do dependent texture lookups, calculate with world scale coordinates etc. Partial precision isn't a one-size-fits-all type of thing really.

[DeanoC]

DeanoC

I do not totally agree with some of your words. I admit that single precision can save silicons, but there is one thing, truely important, that the memory bandwidth and the total amount of memory are limited, therefore it is quite obvious for a programmer to think about the memory utilization. As us sofeware programmer we always try to optimize our code to consume as less memory as we can and to make it as faster as we can.

In game programming, multiple precision will save valuable bandwidth and make it for other features or functions and it provides more flexible way for the programmer to optimize their code. You know what, if the precision is fixed on one, then, some years later, more and more programmer will lose their sense in these optimization skills.

Your talking about external precision (the pack/unpack into memory), which all cards support multiple precision fully (you can pretty much chose external precision seperate from internal). This is completely unrelated to internal precision, on some architecture we use 1 bit integer external precision and 32 bit float internal precision.

Games programming gains little by differing internal precisions, the only advantage I can think of is when doing integer maths and relying of overflow/saturate semantics.

External precision is however a different kettle of fish, we can trade memory, bandwidth and cache efficency.

[digitalwanderer]

How do you pronounce "cache" Deano? :|

[DeanoC]

How do you pronounce "cache" Deano? :|

Well with my mixed Cambridge/Suffolk/London accent I pronounce it similar to "cash" but I believe thats wrong and it should be pronouced closer to "catch".

But like I said I don't exactly speak Queen's English to start with, so who knows? Geneally people know I'm talking about processors and not squids (that will probably confuse the non UK based people) :-)

[digitalwanderer]

How do you pronounce "cache" Deano? :|

Well with my mixed Cambridge/Suffolk/London accent I pronounce it similar to "cash" but I believe thats wrong and it should be pronouced closer to "catch".

But like I said I don't exactly speak Queen's English to start with, so who knows? Geneally people know I'm talking about processors and not squids (that will probably confuse the non UK based people) :-)
Sorry, it's just a running debate I have going on a couple of forums and I wanted to ask someone who I felt was "in the know" for their opinion. Thank you.

I still pronounce it "cash-ay", I learned all I know about PCs online and never really heard the terms said out loud....I embarress myself a LOT talking to people on the phone. (Thus one of the many reasons I ignore and avoid the annoying contraption. ;) )

Thanks again, sorry for the interruption.

[Bouncing Zabaglione Bros.]

How do you pronounce "cache" Deano? :|

Well with my mixed Cambridge/Suffolk/London accent I pronounce it similar to "cash" but I believe thats wrong and it should be pronouced closer to "catch".

But like I said I don't exactly speak Queen's English to start with, so who knows? Geneally people know I'm talking about processors and not squids (that will probably confuse the non UK based people) :-)

http://www.m-w.com/cgi-bin/dictionary?book=Dictionary&va=cache&x=0&y=0

pronouced "kash", from the french "cacher".

[radar1200gs]

ca-che
say the first part like you do for "cable"
the second part shh (shush, be quiet).

[Mariner]

Well, my Collins English Dictionary contains a phonetic guide for the words therein and it tells me I am pronouncing it correctly:

Cash or Kash, if you prefer.

Where the Ca is pronounced as it is in the word Cat and the sh is pronounced as if it was in the word Shoe. 8)

[AlStrong]

I always thought it was a french word... must be a bias from all those years learning the language. :wink:

[Scali]

Isn't the 3dmark03 test pretty static?
I thought that CPU impact on 3dmark03 was very very little. No sound no physics no nothing almost :/
Or am i wrong? Can someone clue me in here?

3dmark03 has a number of dynamic skinned characters (space troopers, trolls, lady with sword). Should be quite similar to Doom3 actually. The physics are also computed in realtime, with Havok, I believe.
CPU impact on 3dmark03 is indeed very little, since the GPU does all the skinning and shadowing, while Doom3 doesn't. Which was my point. Perhaps if Doom3 offloaded more of the geometry processing to the GPU, like 3dmark03 did, it would have a lower CPU requirement, while having the same graphics performance.
That is what I'm wondering... was 3dmark03 right after all?

[Dave Baumann]

That is what I'm wondering... was 3dmark03 right after all?

After the NVIDIA whitepaper went out talking about this - saying it wasn't the "right" thing to do, I remember seeing numerous discussions about it, including an interesting one over at opengl.org. Basically there were opinions on either side and I guess it just comes down to programmer preference and what they feel their engine/game utilisation will be like on the hardware that its going to actually be run on when its available.

I believe the issue with the 3DMark03 method was that doing the skinning on the graphics side meant that there were more passes needed, which would then overly burden the graphics (although I believe those passes could be reduced with later VS models).

[Scali]

I believe the issue with the 3DMark03 method was that doing the skinning on the graphics side meant that there were more passes needed, which would then overly burden the graphics (although I believe those passes could be reduced with later VS models).

Hm, I'm not sure about that. I do recall that they used ps1.4 in gt2/3 because it could reduce the number of passes, and NVIDIA wasn't very happy about that, because their FX cards weren't out yet, so all their cards had to run with ps1.1.
Other than that, I thought the entire gt2/gt3 tests were DX8 level, so they would have used vs1.1, and not vs2.0.

Anyway, that isn't really interesting... The NVIDIA whitepaper was, though... But again, that probably has to do with the fact that ATi had better vertexprocessing options. Even their FX only had 3 units, while ATi had 4. Makes me wonder though... Did NVIDIA pressure Carmack into using their preferred approach, because of the FX series?

Personally I prefer the 3dmark03 approach, because it scales better with high polycount, and is more future-proof (performance of GPUs increases much faster than that of CPUs).
And when I see Doom3, I see high CPU requirements and low polycount...
Not sure if they're related, but I'll stick to my choice for now :)

[Cryect]

How many people still use 32bit integer for boolean variables? Yes, 32 bit is higher precision, but hell, that's many bits for 1 and 0.

Heh, how about all those people who use bool :P

[Dave Baumann]

Hm, I'm not sure about that. I do recall that they used ps1.4 in gt2/3 because it could reduce the number of passes, and NVIDIA wasn't very happy about that, because their FX cards weren't out yet, so all their cards had to run with ps1.1.
Other than that, I thought the entire gt2/gt3 tests were DX8 level, so they would have used vs1.1, and not vs2.0.

Actually, evidently they appear to be complaining about the number of passes required in the VS, claiming it to be VS limited:

3DMark03 uses an approach that adds six times the number of vertices required for the extrusion. In our five light example, this is the equivalent of skinning each object 36 times! No game would ever do this. This approach creates such a serious bottleneck in the vertex portion of the graphics pipeline that the remainder of the graphics engine (texturing, pixel programs, raster operations, etc.) never gets an opportunity to stretch its legs.

It has been alleged, that there is a design flaw in the code related to skinning the characters multiple times with the hardware vertex shader. It is an erroneous allegation and an invalid argument based on the following facts. 3DMark03 is designed to measure DirectX 9 compatible hardware. This level of hardware does skinning several times faster than the CPU. This can be confirmed using the vertex shader test of 3DMark03, which is designed to measure above all skinning speed. CPU vs. vertex shader skinning can easily be compared by running this test with and without software forced vertex shaders. For example, a DirectX 9 graphics card and high-end CPU (ATI Radeon 9700 Pro and Intel Pentium4 3 GHz) gets five times lower frame rates with CPU skinning than with hardware accelerated vertex skinning. An older CPU (Intel PentiumIII 800 MHz) skins more than 20 times slower on the CPU than with the hardware acceleration. Only first generation DirectX 8 hardware is likely to benefit from CPU skinning, since the vertex shader performance advantage of those cards, as compared to the CPU skinning speed, is smaller. Then again, first generation DirectX 8 hardware should rather be benchmarked using 3DMark2001 SE. Since each light is performance-wise expensive, game developers have level designs optimized so that as few lights as possible are used concurrently on one character. Following this practice, 3DMark03 sometimes uses as many as two lights that reach a character concurrently, not five as mentioned in some instances. Thus, vertex shader skinning will be more efficient than CPU skinning on well-programmed DirectX 9 games on DirectX 9 level of hardware. We believe that rational game developers will opt for vertex shader skinning in forthcoming DirectX 9 games.

Does this mean that 3DMark03 is now completely bottlenecked by the vertex shader performance? No it does not. Try running 3DMark03 in different resolutions. If the benchmark was vertex shader limited, you would get the same score on all runs, since the amount of vertex shader work remains the same despite the resolution change. Game tests 2 and 3 scale very well with the resolution, and are thereby mostly pixel shader limited. Changing the skinning to the CPU would reduce the vertex shader workload, making 3DMark even more bottlenecked by pixel shader performance. We did some simple experiments with this rendering technique and CPU vs. vertex shader skinning. The test results did not change much at all, the overall performance only dropped somewhat using CPU skinning. This was to be expected, looking at the difference in skinning speed between the CPU and the hardware vertex shader.

So, in general, I think that if they had higher VS performance they probably wouldn't have made an issue out of the point.

[jvd]

I don't understand why people will kick a big fuss over PP. In every day programming, we save bits where possible, and if sufficient precision is achieved using fewer bits, do it.

How many people still use 32bit integer for boolean variables? Yes, 32 bit is higher precision, but hell, that's many bits for 1 and 0.

Likewise, if JC thinks he has enough bits for his engine, I think nobody has any grounds for criticism. But should PP prove to produce precision errors and cause visual effects, yes, we can call JC out on this.

I'm in a puff because some people are claiming there is no speed diffrence between full and partial. Which is not true if nvidia is asking devs and forcing it through drivers to get partial percision or less.

[Scali]

Actually, evidently they appear to be complaining about the number of passes required in the VS, claiming it to be VS limited

Ah right, if you are talking about the amount of work... I always associate 'passes' with 'renderpasses'. Although even then, NVIDIA has a point... As I said before, their cards at the time had to run the ps1.1 path, requiring extra passes. This would also require the skinning to be executed for every pass, meaning more vertexprocessing overhead.
Then again, that was exactly the point for ATi to introduce ps1.4, and for FutureMark and various games to use ps1.4, it reduces workload and improves performance and/or quality :)

But that was a different era. GeForce4 is now long forgotten... Which makes me wonder why Doom3 still uses that approach (that is, assuming it still uses that approach, Carmack never mentioned a change to the approach afaik), which should now be outdated.
As FM already says in its reply, only first-generation DX8 hardware would benefit from the approach. Any modern GPU easily outperforms the average CPU, so the extra work is justified. They speak of 9700Pro vs P4 3 GHz, and a factor 5 difference. What about the currently fastest... X800XT vs P4 3.6 GHz?
I bet the factor has grown to at least 10 times.

[Farid]

I always thought it was a french word... must be a bias from all those years learning the language. :wink:

Indeed

[French, from cacher, to hide, from Old French, to press, hide, from Vulgar Latin *cocticre, to store, pack together, frequentative of Latin coctre, to constrain, from coctus, past participle of cgere, to force. See cogent]

So cache in english have to be pronounced "cash"...

Great Topic diversion Digi... :lol:

[digitalwanderer]

Great Topic diversion Digi... :lol:
Thanks. It's an old one, but it never fails. 8)

[Nick Spolec]

Hmm.

I've always pronounced it "cahch" (or kah-ch).

But then again, I'm not good with pronunciation. I've always pronounced Geforce as Gayforce.

[AlStrong]

:lol:

[digitalwanderer]

I've always pronounced it "cahch" (or kah-ch).
Is it two syllables the way you do it? Or like one with a sneeze or something in it? :|

[Smurfie]

I'm in a puff because some people are claiming there is no speed diffrence between full and partial. Which is not true if nvidia is asking devs and forcing it through drivers to get partial percision or less.

I really still don't see why you should get all puffed up. It's about having enough bits to do what's required. If Nvidia or the developers in the world feel that 16 bits is enough, that's their choice. So long as no precision is compromised using 16 bits, that is fine. Of course, if precision is compromised, then take issue with Nvidia or the developers.

There's a good reason why ATI went 24 bits. Simply because it's enough. It makes sense to save on the silicon and gain performance, for sufficient precision. Why go 32 bits when 24 bits is sufficient? When 24 bits becomes insufficient, then go 32 bits.

50 bucks is better than 20 bucks, right? But if you are buying a bar of candy, why carry 50 bucks to the store? Simple analogy.

ps. Does anyone have a good writeup on ATI's 24 bit design decision? I think it might be a good read. It's just interesting to see a deviation from the "normal" 16 and 32 bits.

[Razor1]

ps. Does anyone have a good writeup on ATI's 24 bit design decision? I think it might be a good read. It's just interesting to see a deviation from the "normal" 16 and 32 bits.

That write up would be nice, but I think ATI keeps that kind of information under wraps.

[Dave Baumann]

Search these forums - it has more or less been explained by them.

[Reverend]

ps. Does anyone have a good writeup on ATI's 24 bit design decision? I think it might be a good read. It's just interesting to see a deviation from the "normal" 16 and 32 bits.

That write up would be nice, but I think ATI keeps that kind of information under wraps.
http://www.beyond3d.com/forum/viewtopic.php?p=125542#125542

The topic, which is one I raised, resulted in communications (within the thread) between me and sireric, a rather important ATI employee, about DX9, IEEE-32 and ATI's 24-bit. Ignore the personal debates, focus on the exchanges between me and sireric.

That thread should suffice as the "write up" you want, I think. Basically, ATI made their 24-bit decision based mostly on costs, DX9 spec adherence and evolution of the industry. And probably most importantly is that ATI "anticipated" (it's a big word in the context of what you asked) that their 24-bit will meet MS/DX9's "expected spec" during the timeframe of the release of DX9 and ATI's R300 (the/their first "DX9 hardware").

Hope this helps.

[Razor1]

Thx Rev :)

[nelg]

ps. Does anyone have a good writeup on ATI's 24 bit design decision? I think it might be a good read. It's just interesting to see a deviation from the "normal" 16 and 32 bits.
http://www.beyond3d.com/forum/viewtopic.php?t=6223&postdays=0&postorder=asc&

[Nick Spolec]

I've always pronounced it "cahch" (or kah-ch).
Is it two syllables the way you do it? Or like one with a sneeze or something in it? :|

Just one word. As if it were spelled "cach", without the e at the end. I seem to forget that with the e at the end, it should be pronounced differently.

[Smurfie]

[quote="Reverendhttp://www.beyond3d.com/forum/viewtopic.php?p=125542#125542

Hope this helps.[/quote]

Thanks Rev!

[digitalwanderer]

Just one word. As if it were spelled "cach", without the e at the end. I seem to forget that with the e at the end, it should be pronounced differently.
It keeps coming out "kosh" for me when I pronounce it that way. :?

I'm sticking with 'cash-ay', it just makes more sense to me and I reckon I'm a bit too old to give much of a rat's ass if people think I'm ignorant. ;)

[Evildeus]

Well it's ka-ssshhh at least in french ;)

[Reverend]

This thread is close to getting locked unless you guys know how to pronounce "o-n t-o-p-i-c". One or two more OT posts will have this thread locked.

[digitalwanderer]

This thread is close to getting locked unless you guys know how to pronounce "o-n t-o-p-i-c". One or two more OT posts will have this thread locked.
That's easy, I pronounce it "on-top-ick"... ;)

:lol: I'll stop, my apologies! :oops: It's just the house-hunting stressing me out. :(

[Nick Spolec]

Just one word. As if it were spelled "cach", without the e at the end. I seem to forget that with the e at the end, it should be pronounced differently.
It keeps coming out "kosh" for me when I pronounce it that way. :?

I'm sticking with 'cash-ay', it just makes more sense to me and I reckon I'm a bit too old to give much of a rat's ass if people think I'm ignorant. ;)

Think "Ed Koch" (Former NY Mayor).

P.S. Sorry Rev :)

[Reverend]

You guys don't listen. And you guys complain at the same time.

Don't apologize to me.

Locked (and all of you can blame the two that posted after my post above).