FX5900 vs Radeon9800Pro

Introduction

The R300 chip (ATI Radeon9700) has become the world's first graphics processor to offer full support for DirectX 9 and to take a lead over its age-old competitor during half a year. What's more, when NV30 (FX5800) finally arrived to the reviewers (while ATI Radeon9700 had been on sale in every computer shop by then) it turned out that by the consumer qualities it wouldn't be a match for ATI Radeon9700. It was really expensive, lost at performance, the driver hadn't been optimized enough for the entirely new FX micro-architecture by then, and the main thing was video cards on its base were heating up and whistling too much.

Nevertheless, nVidia's partnering companies were unanimously demonstrating futurist masterpieces made with NV30 at expos but were strangely reluctant to sell them. The main cause is the card proved too expensive, and the tradeoff for using scarce DDR-II memory was prohibitive. Even on the condition that nVidia had to sell NV30 chips at a loss! As they say, the prestige is above any price.

The miseries were lasting for half a year. They were trying to switch NV30 to the professional sector Quadro where prices as high as $1000 is a norm, with NV35 (FX5900) declared as a new king of the hill. In response to that, ATI released two new Radeon9800Pro video cards based on the R350 chip.

The current stage of competition between ATI and nVidia looks like this:

Level	nVidia	ATI
Hi-End	FX5900 (NV35)	Radeon 9800
Middle	FX5600 (NV31)	Radeon 9600
Low-End	FX5200 (NV34)	Radeon 9200

The inquisitive reader might easily point to the lack of NV35 on sale, while Radeon9800Pro has been on sale for a month. That's true, but let's be indulgent. If NV35 has been sent out to testers, it means the video card does exist and let's hope it will not suffer the fate of its predecessor, and the production in sufficient quantities will be arranged anyway.

Gigabyte 9800Pro versus ASUS V9950

Today we'll be taking a closer look at both rivals. ASUS' V9950 128Mb (GF5900) has been called up to uphold the honor of nVidia. For ATI, two versions of R9800Pro - 128 and 256 MB - made by Gigabyte come into play. These cards are in a way "reference boards" since their core speed values fully match those recommended.

VGA	Gigabyte RADEON 9800 PRO	Asus V9950
Core	R350	NV35
Process technology, (mk)	0.15	0.13
Q-ty of transistors	110 mln	130 mln
Core speed	380MHz	400MHz
Pixel pipelines	8	8 (or 4)
Fill Rate (Mpixels/s)	3,040	3,600 or (1,800)
Texture units per pipeline	1	1 (or 2)
Textures processed per Texture Unit	8 to 16	8 to 16
Memory speed	340MHz (680MHz DDR)	425MHz (850MHz DDR)
Memory bus DDR	256-bit	256-bit
Memory bandwidth	21.8GB/s	27.2GB/s
Proprietary technologies	SMARTSHADER 2.1 SMOOTHVISION 2.1 - Color Compression (up to 6:1 ratio HYPER Z III + : Z-Compression (up to 24:1 ratio) and Z-cache optimized for real-time shadow rendering TRUFORM 2.0	CineFX 2.0 Engine LMA II Intellisample HCT Technology - Color and Z-Compression (up to 4:1 ratio)
Pixel shaders	2.0	2.0+
Vertex shaders	2.0+	2.0+
FSAA mode (max)	up to 6x	up to 8x
AF mode (max)	up to 16x	up to 8x
Support for DirectX	DirectX 9.0
Additional features	Dual integrated 10-bit per channel 400 MHz DACs Integrated TV Output support up to 1024x768 resolution Integrated 165 MHz TMDS transmitter (DVI 1.0 compliant) VideoShader HydraVision	Dual integrated 400 MHz DACs Integrated TV Output support up to 1024x768 resolution Video Mixing Renderer (VMR) Extended programmability with Cg nView DVC 3.0

Distinctions between R350 and R300

Having seen a lot of nVidia's miseries in migrating to the 0.13 mk process technology, ATI has put off its technology breakthrough and stayed as before at the 0.15 bar.

Basically, the 0.13 mk process allows reducing the die area and place more active elements on the wafer, reduce the power voltage and thus power consumption with heat emission. But that is only in theory. In practice, all is quite the contrary - the coarse and expensive process technology sharply reduces the chip yield ratio per wafer, but nVidia has failed to reduce the core voltage, so in the end NV30 video cards held a new record for the industry at heat emission levels and noise, whereas ATI even in its latest cards recommends to use absolutely standard coolers.

Gigabyte Radeon 9800Pro 256Mb

So, ATI stayed at 0.15, and what's curious, in the R350 chip (9800Pro) as compared to the R300 ( 9700Pro) top model the number of active elements (transistors) has also not gone up but stayed at 110 mln as before.

So what's new in the R350 chip as compared to R300? Putting it straight, not much.

• The core speed has succeeded to be raised from previous 325 MHz (in R300) to 380 MHz (in R350), and the memory operating frequency has been pushed up from 310 MHz to 340 MHz, which gave a rise for the peak memory bandwidth to 21, 8 GB/s (versus 16.0 GB/s in NV30).
• SmartShader 2.1 - the F-buffer has been added. In fact, the competitor to nVidia's CineFX engine allows executing the code of shader programs of unlimited length.
• SmoothVision 2.1 - the quality of FSAA & AF execution has been improved.
• HyperZ III+ - optimized Z-buffer cache which is now tuned for better shader processing.

These changes can be frankly regarded as "cosmetic" but as the benchmarking will show, this ordinary overclocking has sufficed to withstand decent competition with NV35.

Distinctions between NV30 and NV35

At that, things are more comical. The only thing that has undergone alteration is the memory interface. Expensive and scarce DDR-II with narrower 128-bit memory bus used in NV30 has been replaced with more accessible DDR that offers the 256-bit bus, which allowed raising the memory bandwidth to 25-27 GB/s depending on the frequency (variable between 400 and 450 MHz). No doubt the move is absolutely correct. In the design terms, this allows getting rid of huge cooling systems required for DDR-II running at 500 MHz as well as today's 400 MHz - they are not rare and they also need to be cooled.

Another matter is that manufacturers skilled at research of cooling systems in developing NV30 cards came across this task already fully armed and were able to create virtually noiseless cards.

ASUS V9950 128Mb (GF5900)

Preliminary findings on the chips:

Neither R350 nor NV35 are "revolutionary" solutions. The revolution occurred just before the release of 9700Pro(R300) and FX5800(NV30).

In the end, R350 proved to be a modification to R300, and NV35 is a brushed up version of NV30 chip, first in the FX line. Calling NV30 "unsuccessful", we mean primarily its consumer qualities like price, noise, bulk and temperatures. The first, second and third didn't fit within any bounds (there aren't many amateurs of the extreme in this world, nor they have the key role), and the performance wasn't worth such expensive tricks.

It's impossible to release new architecture of the graphic chip every 6 months. Nor does it make sense taking this 'modification' as a time of erroneous expectations. This is more likely a time of releasing mature and interesting solutions where all the mistakes made (inevitable in revolutions) have been taken into account, and the programmers at both companies had enough time to optimize the driver compilers for better performance.

Benchmarking

All the video cards arrived with the most up-to-date official drivers pre-installed. To date, these are Detonator 44.03 for all the nVidia cards and Catalyst 3.4 for all ATI's cards of the Radeon line. Those were just these drivers recorded on the bundled CDs.

What's nice, Asus started marking the driver version straight on the disk surface. Formerly, you had to insert the disk to find out the info. A very good initiative - why not to add it to the armory for all the manufacturers!

Today's unofficial version are Detonator 44,45 and Catalyst3,5. But unofficial drivers merit a special review.

To test products of this class, we've got to proceed from the two prerequisites. First, cards at $400-500 are never fitted into a weak PC, and secondly - on the market there aren't any solutions able revealing their potentials to the full. Both FX5900 and Radeon9800Pro are "future-proof" video cards. Even Quake 3 shows incredibly high fps rate over 500 fps at default settings, and these results produced with outdated benchmarks can no longer be trusted. Game developers lag behind hardware manufacturers. Two more years of the lag and consumers might cease striving for novelties - there won't be solutions to give enough load to the hardware. Intel and AMD suffer from this state of affairs to the full either.

The first new-generation game will certainly be the creature of ID Software - Doom3. But it's still half a year until it is released. But the demo currently on testers' hands is still very raw. The guys are creating models, working out the plot, drawing the levels and ignore the optimization - what matters is that it simply works. Concurrently, optimization for all the VPUs existing on the market is under way, so no one except ID Software could tell at what point all these works are progressing and which VPU is better. Once the release or at least the official demo is out, then we can run tests.

Test bed:

CPU	P4 2.4Mhz (Northwood Ñ1)
Motherboard	Epox 4PDA (i865PE)
Memory	PC2700 2x256MB in the dual-channel mode Timings set to 2:6:3:3
Video cards	Asus V9950 128 Mb (FX5900) Gigabyte Radeon9800Pro 128Mb Gigabyte Radeon9800Pro 256Mb ATI Radeon 9700Pro 128Mb MSI Ti4800SE-8x 128Mb
OS	WinXP + SP1
Drivers	Detonator 44.03 Catalyst 3.4

Benchmarking software used:

1. 3DMark2003
2. 3DMark2001SE
3. Comanche 4 Bench
4. Codecreatures Benchmark Pro
5. Village Mark
6. RightMark Video Analyzer v0.4
7. Unreal Tournament 2003

We did a series of synthetic and gaming tests with benchmarks able somehow to provide enough load to ATI 9800 PRO è FX5900. All the tests were conducted with the 4xAA/8xAF options enabled. The weaker ATI 9700 PRO and Ti4800 were also posed into the same hard conditions. Mind you, it doesn't make any sense to buy a video card to enjoy the "ladders" and jagged lines. Gaming video cards of the top price sector are aimed exclusively at implementing the highest image quality at high resolutions with the good gameplay preserved.

Let me put it straight in advance - the results are not going to be high. We are not pursuing ultimate score points as is the case with testing CPUs and motherboards - we simply want to load these monsters to the full and see their worth.

Notes:

Why Ti4800 was taken for comparisons? Not a previous generation card on NV30, but NV28? Just for your awareness, the NV30 hasn't gained a wide distribution and users almost don't have it on hands, but Ti4200-4800 have been sold in huge quantities. Their owners will better see the difference in results between what they already have and what is offered..

Another note - in fact, we tested TWO VERSIONS of 9800PRO - with 128 and 256 MB both made by Gigabyte. But the cards have shown ABSOLUTELY IDENTICAL figures in all the tests despite the differences in 350 MHz memory speeds for the 256 MB make and 340 MHz for the 128 MB version. We were first puzzled by the identical results - this simply can't be that way! To spend $400 for a 128 MB card or $500 for a 256 MB and see no difference in any of the tests.. The clue was behind the memory organization - Gigabyte had to double the number of memory chips and all the extra was all eaten up by the doubled wiring and load. Another conclusion that comes out of this fact is this - modern gaming applications and benchmarks do not load up even 128 MB of the video memory.

Third note: you won't notice any difference in 3D image quality in real-world applications. Only using specialized test images and their further analysis with Photoshop allows to detect different positions of some pixels which make up the image. By now, the hot debates on the need for using the 32-bit color mode have died down quite a long time ago (in 3DMark2003, the 16-bit mode has been removed) and to date it's absolutely absurd to compare "image quality with the FSAA&AA enabled". The image quality is fantastic in games either, and there's no visible difference between R350 and NV35.

3DMark2003 ver230

This is a synthetic benchmark offering full support for DX9 To start with, the overall score:

Don't get confused by the low overall score points attained because 4xAA/8xAF is a serious trial. As we see, NV35 takes a lead at low resolutions and a little bit lags behind R350 at 1600x1200. But Catalyst 3.4 easily left 9800Pro behind at 2048x1536x32, while neither Detonator 44.03 nor Detonator 44.45 coped with that resolution - three attempts to pass 3DMark2003 at 2048x1536x32 for NV35 ended up in system failure. It's just the driver to blame for that - there is nothing else to explain it by, then both the 128 MB version of Gigabyte 9800Pro also ran the test smoothly as its 256 MB make.

Having summed up the benchmark results, we get a more evident picture:

Unlike all the other cards that took part in the tests, the Ti4800-8x (NV28) chip does not support DX9, so its serious failure was caused by that it simply skipped the fourth benchmark (Mother Nature) in 3DMark2003. and it's just this benchmarks that adds maximum score points.

3DMark2001 SE Pro

NV35 demonstrates a very strange curve - the drops in the low and high resolutions combined with absolute identity of results for the middle resolutions as compared to R350. In the end, ATI is a leader in the total graph.

In the equitable conditions of 3DMark2001 SE Pro where the advantages of DX9 don't count, the Ti4800-8x chips doesn't look so terribly outdated. And that's absolutely right! Until mass quantities of new games supporting DX9 appear - it's just 3DMark2001 that reflects a more realistic difference of the used cards. By the way, the widespread fears in that DX9 games will require mandatory migration to new cards to suffice the start-up have no grounds. No one has yet abolished the software rendering although it is less productive and results in serious torture of a load for the CPU.

Comanche 4 Bench 4xAA/8xAF

This test can be regarded as outdated and simple (but there aren't any new, aren't there?), but it is made on a real engine based on the super-popular game.

What to say about that? Only one thing - with the high quality settings enabled, all the new video cards remain usable for games up to 1600x1200! At the same time, Ti4800 (a performance favorite of the last year) stays playable at no more than 800x600.

Codecreatures Benchmark Pro

It offers a next-generation engine. Gives a real torture of a load, even the AA had to be disabled. At this test, none of the last-year cards was able to show decent gaming FPS rate, while for modern cards getting to the incredibly high resolutions is a piece of cake! ATI 9800Pro at this test left the competitor well behind.

Village Mark

The nice thing about this test is that it was made by a third-party competitor - PowerVR, manufacturer of graphic chips :-)

That's where the only pleasant thing is over ... the test is terribly primitive - the doors and windows look as if drawn on cardboard walls of shabby village cottages rather than "exist".

As is the case with the Codecreatures Benchmark Pro, the performance of NV35 is right between R350 and R300.

Unreal Tournament 2003 Demo

That's a popular graphics test based on the Unreal Engine built on DirectX7 with speckles of some DirectX8 functions.

For more details of the test and the UT2K3 utility, read the review "Unreal Tournament 2003 as a benchmark"

The test is anyway a bit old-fashioned for today's video cards albeit looks good. It's completely evident that at low resolutions the cards come up against the performance of the engine itself. R350 rules as usual at high resolutions - 58 fps vs. 42 fps for NV35. At more standard 1280x1024, the cards demonstrate approximately the same and even very game-able results. And that under 4xAA/8xAF!

Note that Ti4800 at these quality settings remains gameable at only 640x480.

It makes a big sense for UT fans to buy something modern, set all the options to the maximum and enjoy the thrilling graphics with the gameplay dynamics preserved.

That is only in the Direct3D mode, and in using OpenGL we see a serious failure for nVidia cards.

The same picture is seen in all the other built-in demos. A fight between R350 and NV35 on par in the Direct3D mode, with a complete fiasco for nVidia cards in the OpenGL. It's a mystery what's behind that.

The results for NV28 in OpenGL do not lag much behind those for NV35. Seems like Detonator developers are still unable to fully optimize the driver for the new FX architecture in the OpenGL mode and have thrown in all efforts to polishing the Direct3D. But then, using FX in professional packages appears to be quite problematic.

RightMark Video Analyzer

RightMark Video Analyzer is a fancier of ATI cards :) That's a complete fiasco for FX5900 which hardly copes with shader operations. On the other hand, there aren't any games using shaders so far, so - it's not that awful anyway.

Findings

To start with, here is an instructive story. Since the beginning of opposition of the three graphics monsters - 3dfx, ATI and nVidia, a special karma has been pursuing them for years:

3dfx are terrible snobs too arrogant even towards their fans, absolutely confident in their holiness, right of the first born and everlasting leadership, which played a bad trick to them. For the time being there, these guys have contrived to arrange no more than a single presentation, no party, no seminar and no press conference outside the USA! It was absolutely impossible to persuade them for a meeting... In the end, they changed their ways to the better... After their closing, they gave interviews here and there making up for the past lost years.

ATI is a marvelous team of technical engineers with everlastingly poor drivers and awful marketing. Just fancy their entangled numeration for products and names for endless drivers until Catalyst finally appeared. That's terrible!

nVidia is the youngest, energetic and impudent a team. Their chips hit precisely to the market boiling point, and their unified driver polished to mirror reflection has turned into a symbol of the company now demonstrates exceptional performance and stability. One can't help noticing their correct and aggressive marketing and close relationships with game developers. Since the day of their inception, their slogan has been "we'll do it!".

• the erroneous stake made on DDR-II, but we have already mentioned it and won't repeat again. Already fixed..
• and secondly the new architecture of the chip implied creating a compiler for the new chip from scratch. The depleted team of Nick Tiantos failed to cope with the posed task in good time. It depleted due to the only simple reason - the departure of some employees with further establishment of an independent team of developers which concluded the only contract. Guess with whom! With ATI! It's just this team that has developed Catalyst, a unified driver for ATI.

One lagged behind, another caught up with the rest - it turns out both companies to date have approximately equal potentials. From the consumer's viewpoint, this outcome is the best. The lack of sure favorite makes shifting the competition to the price range.

Assuming the chosen cards absolutely precisely reflect the essence of opposition between R350 and NV35, things look like this:

- In most tests that we conducted, ATI 9800Pro demonstrates advantages over FX5900. Let's list them: RightMark Video Analyzer, Unreal Tournament 2003, Village Mark, Codecreatures Benchmark Pro, Comanche 4 Bench, 3DMark2001 SE Pro. In fact, the only test where nVidia takes a lead over ATI was 3DMark2003. But that's true provided the FSAA functions are enabled in the drivers; if disabled, things will look like this:

Isn't it a curious graph? :) At least, it's visually seen what we pay for.

While FX5800 in due time was unable to outrun Radeon9700Pro, this time FX5900 in most tests lags behind Radeon9800Pro a little bit. But nVidia has the Ultra option in stock and things may change about it. At least, Gainward's FX5900 whose frequencies (440/900) strongly differ from the canonical (425/800Mhz) already demonstrates some advantages and things are not the way they seem. This is to say that frequencies in the Ultra version will be much higher.

In our review we tried to compare just the chips without digressing towards video cards and assuming them as the "reference". But reality is such that every manufacturer experiments with the values of core and memory speed, cooling systems, functionality, let alone the package bundle etc.. In the near future, we'll be doing a detailed round-up of Asus V9950 and Gainward Ultra/1200 (both based on FX5900) which are worth each other.

Big thanks to:

• Asus & IT-Labs - for the sample Asus V9950 video card
• Gigabyte for submitting Gigabyte RADEON 9800 PRO sample;
• Atlantik for presenting a Radeon9700Pro sample.