PDA

View Full Version : 5900ULTRA Review, ouch.


Pages : [1] 2 3

creedamd
06-27-03, 10:17 AM
Good review, seems it gets smacked around abit by the 9800pro 128mb card.

http://www.ocaddiction.com/articles/video/fx_5900ultra_vs_radeon_9800pro/

DivotMaker
06-27-03, 10:27 AM
Not the impression I got at all. It looked like a pretty even match with ATI winning half the fps battles as well as the IQ. However, we all know that can be a subjective thing even though many will say ATI is better and I won't argue with that because it IS subjective and I respect other's opinions.

Uttar
06-27-03, 10:38 AM
BigBerthaEA: Well, IQ *is* subjective, but what is certain is that ATI's AA algorithm is much more complex, and expensive to implement than nVidia's.
So I'd say that since ATI's implementation is widely recognized as successful, ATI's AA IQ is at least equal or better than nVidia's AA IQ.
IMO, ATI's AA IQ is much better than nVidia's - but yes, that part is subjective. The part that it is as good or better, and cannot be worse, is not. All IMO, of course.

Also, I'm only talking AA IQ here. Should we go in AF issues, it all becomes even more complex to discuss, and there frankly, it's all very subjective IMO.


Uttar

DivotMaker
06-27-03, 10:58 AM
Originally posted by Uttar
BigBerthaEA: Well, IQ *is* subjective, but what is certain is that ATI's AA algorithm is much more complex, and expensive to implement than nVidia's.
So I'd say that since ATI's implementation is widely recognized as successful, ATI's AA IQ is at least equal or better than nVidia's AA IQ.
IMO, ATI's AA IQ is much better than nVidia's - but yes, that part is subjective. The part that it is as good or better, and cannot be worse, is not. All IMO, of course.

Also, I'm only talking AA IQ here. Should we go in AF issues, it all becomes even more complex to discuss, and there frankly, it's all very subjective IMO.


Uttar

I agree completely. ATI has some solid advantages versus nVidia and there is no question about it when you look at the screenshots of the UK Gamer review using the AA program.

http://www.ukgamer.com/article.php4?id=5&page=7

Some VERY noticeable differences in the 3 o'clock and 9 o'clock positions as the reviewer noted. This UK Gamer review is one of the most thorough I have ever seen between the two cards and I have a better appreciation of the differences between the two now.

I guess I am so used to playing a game that really does not benefit from AA (Tiger 2003) that I lose sight of other games that DO benefit from AA. Tiger benefits greatly from AF though and that feature is of great interest to me.

Good points...

jbirney
06-27-03, 11:39 AM
but the thing that struck me was how close these two were. Remeber the 5900U has a significant clock advanatage which leads to higher Fill Rates and MB. Once your at high res then its all about brute force. The fact that the R9800 stayed close and won some is amazing. Again it shows you just how good the R300 desgin was and all on the slower .15 um process.....

DivotMaker
06-27-03, 12:16 PM
Originally posted by jbirney
Again it shows you just how good the R300 desgin was and all on the slower .15 um process.....

I have been saying that you really have to hand it to ATI because they flat out have a very good product (9700/9800 Pro) that to me seems to be somewhat of an overachiever...

reever2
06-27-03, 01:21 PM
Originally posted by jbirney
on the slower .15 um process.....

It should be faster .15 um process

poursoul
06-27-03, 01:28 PM
Originally posted by reever2
It should be faster .15 um process

he was just comparing it in his mind to the .13 FX.

Dazz
06-27-03, 02:03 PM
I think ATi made a good move by sticking to well used .15m rather then .13 due to low yeilds. While using basic VPU's on .13m so yeilds are greater :)

jbirney
06-27-03, 02:45 PM
I guess I ment is that we have been told that the .15 um process was at its limits a long time ago. That you really could not make a large die and clock it fast on .15 um. You needed a .13u process to get any speed out of it. NV showed us that they can clock the .13 um very fast on a large IC. BTW I live in Boynton Beach not to far away from you poursoul :)

CaptNKILL
06-27-03, 02:55 PM
Why didnt they use the 256Mb 9800 Pro?

Im sure the results wouldnt be much different, but we are compairing a card that cost around $370 to a card that costs around $500. And the $370 card wins the majority of the tests.

I dont know if they just wanted to make the 5900 Ultra look bad, but It deffinitly makes the 9800 Pro look a LOT better.

EDIT:
Actually, the Radeon can be bought for $340 (http://www.gameve.com/store/gameve_viewitem.asp?idproduct=1089&showit=1) or $345 (http://www.buyxtremegear.com/ati9800128o.html) !!!!!!!!!
Where as, the lowest price for the 5900 Ultra is $504 (http://www.buyxtremegear.com/gain4.html).
:eek:

Dazz
06-27-03, 03:21 PM
Well a plain FX5900 will just make it look just as good :D

stncttr908
06-27-03, 04:20 PM
I'm really an IQ buff. I can't stand jaggies which makes playing on my system incredibly painful. For these reasons, I'm really looking towards a 9800 Pro for my next setup, especially when they drop down to the $300-320 range. I'll be the one stopping in the middle of a firefight to look at the nice AA applied to the steel beams overhead. :D

IISquintsII
06-27-03, 05:47 PM
I have to admitt I've been fragged a couple time when I stop to smell the roses :D

StealthHawk
06-27-03, 06:06 PM
Originally posted by Dazz
Well a plain FX5900 will just make it look just as good :D

How so? The plain gfFX5900 is clocked slower than the gfFX5900Ultra.

Sazar
06-27-03, 06:27 PM
Originally posted by StealthHawk
How so? The plain gfFX5900 is clocked slower than the gfFX5900Ultra.

per the grapevine... does the vanilla 5900 not also come with a 128bit memory bus instead of 256 ?

Nutty
06-27-03, 06:40 PM
I very much doubt it, that would be stupid.

From the specs on etailers sites, it says the 5900 nonUltra, uses 256bit RAM. I assume they mean 256bit memory bus, same as the ultra.

Sazar
06-27-03, 06:41 PM
Originally posted by Nutty
I very much doubt it, that would be stupid.

From the specs on etailers sites, it says the 5900 nonUltra, uses 256bit RAM. I assume they mean 256bit memory bus, same as the ultra.

I had heard from the grapevine that there would be lower end 5900's with 128bit memory...

just wanted confirmation :)

Lezmaka
06-27-03, 09:04 PM
Originally posted by Sazar
I had heard from the grapevine that there would be lower end 5900's with 128bit memory...

just wanted confirmation :)

Around the launch, rumors were that there would be a 5900 Ultra, 5900 and 5900 Value. Differences between Ultra and non ultra were ram and clockspeeds. Differences between 5900 and 5900 Value were clockspeeds and possibly 128bit bus but I hadn't heard that very much.

Morrow
06-28-03, 08:42 AM
Originally posted by jbirney
but the thing that struck me was how close these two were. Remeber the 5900U has a significant clock advanatage which leads to higher Fill Rates and MB. Once your at high res then its all about brute force. The fact that the R9800 stayed close and won some is amazing.

What is so amazing? The fact that the 9800 has double the pixel pipelines than the 5900 or the fact that the 9800 has four times the pixel shader units than the 5900 and still only can beat nvidia flagship by a few percent in some tests?

Considering the small GPU clockspeed difference of 70MHz and the big difference in pipelines and shader units, the 9800 should have outscored the 5900 by at least 30-40%. but nada...

You wanna know more about the architectural differences between those two cards? The nv35 has three vertex engines, the 9800 four! The biggest difference however is that the 9800 can render 48 AA samples per clockspeed (8 pipelines & 6 samples per pipeline) but the nv35 is only able to do one third of it (4 pipelines & 4 samples per pipeline)!

You see the big advantages the 9800 has over the 5900 but just merely succeeds in beating the 5900 in some benchmarks. I would never dare to call that amazing but rather sad or disappointing...

Of course are you now asking yourself why nvidia didn't choose to implement more pipelines and more shader units so they could have easily outscored any R3x0 card for the next few months. The reason for this is because the 0.13 micron manufacturing process didn't allow for more pipelines and shader units without requiring liquid nitrogen to cool the chip. Considering this, ATI did indeed make a wise choice when they decided to stay at 0.15 for the R3x0. They have the advantage for now until the technology is ready for a R3x0 type card at 0.13 micron --> nv40.

CaptNKill:
If the reviewer had used the 256mb version of the 9800 Pro, this would not have changed the benchmark results by much since today's cards are not memory size limited but memory bandwidth limited!

DivotMaker
06-28-03, 08:58 AM
Thanks for the post Morrow...I had no idea that the 9800 had so much more "under the hood" than the 5900. I guess the strategic lesson in all of this is that nVidia had the right intentions to move to .13, but the transition must have had some unseen roadblocks (or roadblocks that appeared to be solve-able when they made the .13 process decision) as you mention, or nVidia would likely have stayed with .15 until the .13 process matured more.

Very interesting and thanks!

noko
06-28-03, 10:04 AM
So the R350 has much more under the hood yet the NV35 has more transitors :confused: . What is going on, is the NV35 crippled or something with sections of the chip not being used? Plus if the IQ where matched up, expecially in the AA department the ATI cards would wipe Nvidia cards clear off the scale so to speak. I know sounds like I am an ATI fanboy, but I am utterly impressed with the IQ out of my Raddy and frame rates graduating from a GF3 card.

John Reynolds
06-28-03, 10:11 AM
Originally posted by Morrow
What is so amazing? The fact that the 9800 has double the pixel pipelines than the 5900 or the fact that the 9800 has four times the pixel shader units than the 5900 and still only can beat nvidia flagship by a few percent in some tests?

Considering the small GPU clockspeed difference of 70MHz and the big difference in pipelines and shader units, the 9800 should have outscored the 5900 by at least 30-40%. but nada...

You wanna know more about the architectural differences between those two cards? The nv35 has three vertex engines, the 9800 four! The biggest difference however is that the 9800 can render 48 AA samples per clockspeed (8 pipelines & 6 samples per pipeline) but the nv35 is only able to do one third of it (4 pipelines & 4 samples per pipeline)!

You see the big advantages the 9800 has over the 5900 but just merely succeeds in beating the 5900 in some benchmarks. I would never dare to call that amazing but rather sad or disappointing...

Of course are you now asking yourself why nvidia didn't choose to implement more pipelines and more shader units so they could have easily outscored any R3x0 card for the next few months. The reason for this is because the 0.13 micron manufacturing process didn't allow for more pipelines and shader units without requiring liquid nitrogen to cool the chip. Considering this, ATI did indeed make a wise choice when they decided to stay at 0.15 for the R3x0. They have the advantage for now until the technology is ready for a R3x0 type card at 0.13 micron --> nv40.

CaptNKill:
If the reviewer had used the 256mb version of the 9800 Pro, this would not have changed the benchmark results by much since today's cards are not memory size limited but memory bandwidth limited!

And how many of the games tested actually stress the shaders, pixel or vertex? Answer: they aren't. As for the pipe differences, considering the number of texture units for each chip the 5900 will still have a higher theoretical multi-textured fill rate due to its clock speed. As for multi-sampling AA performance, the 5900 also has a considerable bandwidth advantage, and memory bandwidth is usually the bottleneck for AA performance.

It's amazing to me how a little knowledge can be so wildly spun into erroneous conclusions.

Skuzzy
06-28-03, 12:25 PM
Morrow missed the mark in a number of areas.

130nm transistors dissipate less power than 150nm transistors at the same clock rate. In other words, they run a bit cooler at the same clock rate. Even though th resistance is higher in a 130nm path, the voltage is lower and the path is shorter, which would negate the capacitance and resistance disadvantages.

The heat from the FX line is probably due to poor layout versus actual transistor count. Remember, ATI has a cross-license arrangement with Intel. Intel learned a lot of layout tricks to get thier 150nm parts to the clock rates they did. ATI was a benefactor of this data and have further refined the process.
Layout is everything when working at these design levels. Want to FUBAR a design? Put two transistors on opposite sides of the silicone that are part of a single register. Not pretty.

I attributed some of the heat isses to poor layout, as other things do not add up and lead to this conclusion. NVidia's NV35 simply should not have the transistor count it does, given what it has exposed to the software as compared to ATI's R350.
Part of it would be the multi precision that NV3x supports, but that really does not take that much, in terms of transistor count, as many of the transistors would be shared anyways.
ATI has displacement mapping, which would require more hardware (transistors) than NVidias precision levels would.

NVidia may have intended to have more features, but the design falled and thus parts of the silicone are not enabled. I go against this theory due to the heat. If parts of the silicone were disabled, there would be less heat generated.

I keep coming back to poor layout. It is the only thing that makes sense, and it is not 100%. Could be a combination of several things. It is hard to know.

Maybe they actually built the ram into the GPU to hold all the pixel/vertex instructions. Would be a bit silly, and not really gain them anything as they can only do a few parallel operations in this area. I guess it could be ram, as that would take a truck load of transistors and need a lot of power.

Morrow
06-28-03, 05:25 PM
Originally posted by Skuzzy
130nm transistors dissipate less power than 150nm transistors at the same clock rate. In other words, they run a bit cooler at the same clock rate. Even though th resistance is higher in a 130nm path, the voltage is lower and the path is shorter, which would negate the capacitance and resistance disadvantages.


0.13 micron does not necessary mean less heat dissipation than 0.15 or 0.18 micron. How surprised do you think was AMD when they found out that the initial 0.13 Thoroughbred core had a higher heat output than the bigger 0.18 Palomino core?

In fact normally when a die shrink is done, the number of transistors the core is consisting of, is often increased more than the smaller heat output due to the dieshrink can compensate. Adding more transistors to the nv3x architecture which could have helped competing better with the R3x0, would have delayed the nv30 even more than it currently did... nvidia was heading into a dead end and the fact that TSMC didn't managed to incorporate a special manufacturing process for better heat dissipation made things even worse. Nevertheless the nv35 is now available and is competing on bar with the 9800 Pro.

Of course are there many more factors resulting in more dissipation than those I listed in my previous post such as the number of PCB layers used or the total transistor count but that was not the point of my post concerning the architectural differences of both high-end cards.

John Reynolds:
I also do know that there is more to performance than just the raw number of pipelines, shader units or whatever but since this is not a tech site, going into details wouldn't have helped much. I also aware of the comparing both architectures like I did is not the right way to do since both companies choose two different ways to do the same work, especially considering texture management and shader calculations. Nevertheless the information I gave is accurate and shows that the performance of the nv35 is not nearly as bad as some people want it to be!

I was just trying to explain why the often heared "R300 = intelligent architecture / nv30 = brute force power" statement is actually wrong. People just seem to focus on the extra 70MHz the 5900 has but completely forget the real raw power advantage of the R350!

Fillrate and multi-texturing due to the 4x2 design are indeed an advantage of the 5900 but in all other areas the 9800 should have a big performance lead which is currently however does not have...