Introduction

Was still an introductory shortly once together to take what we have reported in the news, and some certainly known occurring paragraph allows.

It was primarily the inexplicably high fill rates, both the Radeon 9600 XT (called the following only with the chip name RV360) and the Radeon X 800 Pro, we want to abbreviate not named their chip, since not been clarified is how exactly the Radeon X behaves 800 XT, which has four built-in pixel processors as opposed to the Pro,. These were some fill rates tests the theoretical maximum of what mathematically and with never attainable efficiency would have been possible 100%.

Last week we thanked the members of the PCGH for the contributing of some Radeon X 800 Pro benchmarks, without which this story would have been certainly not so interesting, could the new R420 series in the first test but exceptionally well against the new GeForce 6800 Ultra claim – even the mathematically clearly losing X 800 Pro could bring some victories.

Over the last week we tried article about the CineFX architecture, which we want to thank at this point in close collaboration with R. Kornemann, co-author of the PCGH article on the NV40 technology and author of 3D-Center, exclude systematically all other influences.

Process of elimination

ATi provides in some press presentations and more or less official Communiques out, why, its larger number of quad-based pixel processors times apart from, is superior to the R420 of the previous generation.

• larger text urcaches for improved anistropes filters
• programmable cross-bar storage controller
• HierarchichalZ, part of HyperZ HD, this works now also per pixel processor

The “enhanced cache sizes” could refer of course to the GPU as a whole, and even apply – she do not definitely has an effect on the efficiency. The improved, programmable memory controller not convinced so far still constellation 800 Pro/catalyst 4.5 on X. Armed with Marko Dolencs fill rates testers we have reviewed, how much can take advantage of the R420 with 12 pipelines from these improvements. Result: near zero

We have reviewed removing hidden surfaces, which has been improved in the R420 and HyperZ HD now calls himself, as a factor. Is this even the improved buffer size to the larger formats, allow two popular HD(-TV) now the three-tiered Z-buffer up in resolution of 1920 x 1080. In the older chips of the R3xx series 9800 SE, even no hierarchical Z-buffer is concluded that offer chips of the RV series, as well as the partly deaktivierten Radeon 9500 and Radeon here at 1280 x 1024 of the normal resolutions.

< div >

Here, is measured in discarded pixels per clock and block of four of pixel pipelines (a pixel processor), that the Radeon 9800XT must give to the first and most effective level of hierarchical Z buffer in high resolutions in 1600 x 1200.

Real world measurements

We have here the clarity on the much-used game “Unreal Tournament 2003″ focuses among other things also, because it was possible here, including screenshots in identical position, what we did with the so-called “cheat demo” from the archive of 3D-Center, which at the time the reduced Trilinear filtering for more recent drivers for the GeForce FX dyed MipMap level has shown by using.

Here we see the results of the Radeon 9800XT in the pyramid demo of Unreal Tournament 2003. The values are all identified with the help of our benchmark best practices, which can be viewed here. Driver was the recently released catalyst 4.5.

The results surprise little, can but the X 800 Pro in virtually every throw a substantial performance increase compared to its predecessor on the scales.

Here we have the ATiTools of “w1zzard” served, converted to the higher number of pipeline, the X 800 Pro in the same fill rate memory bandwidth of the Radeon 9800 XT to clock. This was achieved with 274.5 MHz GPU clock and 364,5 MHz RAM stroke, resulting in an approach to 99.9% of the real R9800XT fill rate and bandwidth.

Here so apparently Green apples were compared with red apples – all the more astounding, since we used us naturally identical settings.

As mentioned in the news of the May 7, 9600 XT fell on a Radeon the FPS values achieved some quite clearly, if you entered the console command “r_colormiplevels 1″ – at the time established in the Brecourt demo of call of duty.

Unreal tournament offers the “firstcoloredmip 1″-command a similar procedure – which, for example, the reduced Trilinear filtering in many drivers of the GeForce FX series has been shown.

Reduced performance on pretty much the value reached the predecessor chip, where last outside of measurement tolerances here not breaks 9800 XT on same raw performance such as the Radeon, reduces the use of colored mipmaps miraculously.

But exactly this result to expect has been actually, because the architecture of R420 was apparently hardly changed compared to its predecessor.

These results have reflected in a large number of other games. Whether these well, such as games, build on the Quake3 engine via console command allow coloring, or whether the result was – helped externally via 3D Analyzer for this purpose remains the same.

Also measurements without FSAA and AF were not may be unaffected, fell but not as strong, as the least densely operate games on the filling rate limit, such as for example the VillageMark.

Other affected applications were: SpellForce, Aquamark v2. 3, Delta Force Black Hawk down and call of duty. So that should be secured, that this behavior is not application-specifically limited – ATi has so far also rigid turned against a “application-specific optimization”.

Differential images

Thus alerted, we tried to find out how the R420 was able to achieve here such performance benefits – revealed so far but the optical comparisons dyed mip-level ordinary Trilinear filtering. To be honest, our search continued only at this point for alternative policy options, however, we have already anticipated the result of two pages.

< div >

(The images are lossless compressed PNG files)
beat that with each just under 1.3 MB record)

These shots, left the R9800 XT, right the X 800 Pro, is dealing with original screenshots from the cheat demo for DoD in 1024 x 768 x 32 with each 4xAA and over the rTool forced, as a “Full Tri” prestigious, anisotropic filter. These images were in some places (e.g..) (Marble tiles on top and the bottom of the rock in the background) slight differences on.

A difference method, in which we reviewed the color values exactly the bit with a self-penned tool by Ralf Kornemann, we have can by now black recolor the pixels in the respective chips, which are equal on both images.

For a real Trilinear filtering the identical image area may be very small (always there where due to the LOD, the second mipmap with 0 is received %), as you can see filtering in the well-known AF testers on the narrow, full color strip at trilinearer. Also the Radeon shows here perfect filtering 9800 XT in their technical means.

< div >

(Links the X 800 Pro and right the Radeon 9600 XT)

Because we due to the measurable and visible change in the AF fill rate for the forced by “Full tri” AF on the rTool of which had to assume that are no longer a texture days optimization is, that only at the first texture layer trilinear AF applied and seven to bi-linear AF falls back on all other, we after signs of a reduction on the texture filtering trilinear were looking for. A so-called “brilinearer filter” characterized by, that where the BRI band acts, the image must be exactly the same with the pure bilinear filtered image. A reduced trilinear filter the identical area on the differential images becoming but compared to real trilinearer filtering with the strength of the optimization.

Purely for the sake of completeness we have drawn once more the comparison to the Radeon 9600 XT, which also has a trilinear optimization. This seems to be even stronger than that of the Radeon X 800 because of wider identical volume Pro.

The result seems clear: as soon as one looks at the Radeon X 800 as well as the Radeon 9600 XT isn’t exactly on the finger, say, dyed mip-level checks applied filtering, is an optimization of the Trilinear filtering, which greatly saves on fill rate especially in combination with the texture-stage optimization of AF.

Consequences

Was the outcry when nVidias reduced trilinearer filtering already, because without being asked, the share of real trilinearer filtering is been reduced a new level seems here to be achieved in the last year.

A detection(auf_nichts_anderes_lassen_unsere_Tests_zurueckschliessen) and turning off optimize in an optical inspection with coloured MIPS are dyed mip levels a unique Simians – in school it was said at the Erwischtwerden immediately:

< div >

Six, put!

Particularly negative, of the time of this matter affects. Because of the GeForce had offered nVidia for the launch 6800 still a review driver, which allows you to completely disable the trilinear optimization (as far as we are so far informed – unfortunately at the moment no NV40 available is us!).

Aware that test high end cards only with maximum quality, many reviews used also this switch (as well our article appearing in the near future), to get really maximum quality. On the other hand, the well-known optimization on ATi was tried to bypass using known resources (rTool) and represented in the consciousness of the previous operation of this measure, the results as “comparable”.

Unfortunately seem based on this knowledge much so far on the Internet published benchmarks not comparable and thus more or less useless to be.
The evaluation of the new X 800 series (we dare the conclusion that this phenomenon not only on the X 800 Pro is true) and the Radeon 9600 series, must be put probably first on hold, until really comparable benchmarks can be made.

Thanks to this point again for active help, food for thought and subliminal references to: Ralf Kornemann, the PCGH that user “ow” and “Stefan Payne” from the 3D-Center Forum and Sapphire for the X 800 Pro.

Update from the 18.05.2004:
ATi has delivered an official position on this matter, which of course is contrary to our article.If this opinion correctly and not interpret is of course not maintained – 100%, we can our accusation of a genuine “Täuschungsversuches” on the other hand was not to expect that a clear admission of guilt would follow.

Second update of the 18.05.2004:
Please note the the above linked news update.