Testbed and Methods

We tested Matrox Parhelia-512 in the following system:

• AMD Athlon XP 2000+ CPU;
• MSI K7T266 Pro2 v2.0 (VIA KT266A) mainboard;
• 2x256MB PC2700 CL2.5 DDR SDRAM by Crucial;
• Fujitsu MPF3153AH HDD.

We used the following software:

• 1.0.4.231 driver for Matrox Parhelia 128MB Retail graphics card;
• Detonator 30.82 driver for NVIDIA GeForce4 Ti4200 based graphics card for Windows XP;
• 6.13.10.6166 driver (Catalyst 2.2) for ATI RADEON 8500 128MB graphics card for Windows XP;
• Windows XP;
• 3DMark 2001 SE build 330;
• Quake3 Arena v 1.30;
• Serious Sam: The Second Encounter;
• Unreal Tournament 2003 DEMO v.927;
• Codecult Codecreatures benchmark;
• Comanche4 Benchmark.

Benchmarks Settings were adjusted like that:

3DMark 2001 SE:

We set 32bit frame buffer; 32bit textures, 32bit (24bit) Z-buffer, D3D Hardware T&L / Pure Hardware T&L.

For High Polygon Count and Vertex Shader Speed tests - 640x480, 16bit textures, 16bit frame buffer, 16bit Z-buffer.

For Fill Rate tests - 32bit frame buffer / 32bit textures / 24bit Z-buffer and 16bit frame buffer / 16bit textures / 16bit Z-buffer.

Quake3 Arena:

32bit screen and textures color depth. Maximum graphics quality settings. Tri-linear filtering and texture compression enabled.

Serious Sam: The Second Encounter:

Quality mode: 32bit screen color depth. "Quality" graphics quality settings.

Speed mode: 32bit screen color depth. "Speed" graphics quality settings.

Unreal Tournament 2003 DEMO v.927:

Default graphics quality settings.

Codecult Codecreatures Benchmark:

Anti-Aliasing - off, sound - off, all other settings - default.

Comanche4:

Default graphics quality settings.

Performance

Synthetic Benchmarks

Let's start with the synthetic 3Dmark 2001 SE test set to check the texture rendering speed:

In case of single texture rendering, Matrox Parhelia shows nothing exceptional due to its lower GPU clock-rate - 220MHz against 250MHz in NVIDIA GeForce4 Ti4200 and 275MHz in ATI RADEON 8500.

But during multi-texturing it surpasses its rivals as Parhelia-512 has four texturing units per pipeline while other GPUs - only two.

When textures and frame buffer color depths are set to 32bit, the real texturing speed of Parhelia-512 is only 74% of the theoretical maximum in case of single texture (this maximum is equal to 880Mtexel/sec) and 72% for multi-texturing (the maximum makes 3520Mtexel/sec).

NVIDIA GeForce4 Ti4200 in 32-bit color mode shows 65% of the theoretical maximum for single texture processing and 82% for the multi-texturing. ATI RADEON 8500 notched 72% and 88%, respectively.

Let's see what provokes the biggest performance loss when switching to 32bit color: screen, textures or Z-buffer color depth increase:

In case of single texture rendering, the increase in the frame buffer depth from 16 to 32bit is responsible for the greatest performance loss. Reads and writes into the frame buffer are constantly alternating in this mode, so the result is not at all surprising.

During multi-texturing, the biggest slowdown was provoked by the textures color depth change. It suggests that texture caches of Matrox Parhelia work less efficiently in the multi-texturing mode than in the single texturing one.

It seems that the results in the Fill Rate tests were limited by the inefficient communication between the GPU and the graphics memory, although Parhelia features 256-bit memory bus. Other cards use special technologies ensuring efficient utilization of the available graphics memory bandwidth. They are LMA II in NVIDIA GeForce4 Ti4200 or HyperZ II in ATI RADEON 8500. Parhelia-512 cannot boast any. Also it owes its failure to the low efficiency of texture caches.

Let's prove this statement with the help of VillageMark, the test that displays a scene with a high Overdraw parameter:

The result seems to fully conform to the supposition that Parhelia-512 has no technologies for efficient graphics memory bandwidth utilization.

But this test uses tri-linear filtering and in the case of Matrox Parhelia it means that two texturing units process every texture. So, Parhelia is at even here with its rivals. Namely, the rendering of three textures in this test is split into two steps: the first two textures are rendered first and the third one - during the next step.

Matrox Parhelia was behind NVIDIA GeForce4 Ti4200 and ATI RADEON 8500 in all modes, but the gap becomes smaller in higher resolutions thanks to the 256bit memory bus. Overall, it's clear that Parhelia-512 cannot use up the available graphics memory bandwidth as efficiently as its competitors do.

Parhelia performs this pixel shader much faster than other cards, although the Parhelia-512 chip works at lower core frequency. Well, it could have been wrong to expect anything else from the chip that has four texturing units per pipeline and a 256bit memory bus.

In this test NVIDIA GeForce4 and Matrox Parhelia had to render a texture in two passes, and Parhelia proved the leader here.

According to the official information, Parhelia-512 features four T&L and vertex shaders units that are working in parallel. If that's true, then we can only wonder at their low efficiency: Matrox Parhelia overcame its rivals only in the test with eight light sources. Well, it may be that only one or two vertex shaders units are now enabled in Parhelia-512 for some purpose.

During the software calculations of transformation coordinates and lightning, that is by the CPU, Matrox fell far behind the rivals. Seems like the transfer of the processed vertexes to the GPU is less efficient than in NVIDIA GeForce4 Ti 4200 or ATI RADEON 8500.

The tests show that Parhelia-512 is just slightly faster when processing vertex shaders. It proves that either its vertex shaders units are not that efficient or that some of them are disabled.

Unlike its competitors, Matrox Parhelia-512 has no hardware point sprites support and emulates every sprite with two triangles. So the poor result in this test is quite natural.

Well, the synthetic tests shape up into the following picture: the Matrox Parhelia-512 GPU specs indicate high potential of the chip, but some flaws in the architecture lead to surprisingly poor performance. Let's hope that our hero will prove better in gaming tests. The 256bit memory bus must contribute to this.

3D Games

Matrox Parhelia is an outsider. The "Low details" mode of the test requires no more than two textures to be rendered, so Parhelia cannot use its potential advantage of four texturing units per pipeline (if there had been more textures to be rendered, the rivals would have to spend one more GPU clock). The defeat was caused by the already mentioned fact that Parhelia-512 cannot use its memory bandwidth effectively enough.

Here Matrox Parhelia is on the losing side again. The advantage of four texturing units per pipeline is not called for here. Besides, the character animation is done here via vertex shaders, and Parhelia-512 cannot boast vertex shaders processing much faster than that implemented in NVIDIA GeForce4 Ti4200 and ATI RADEON 8500.

The twice as fast memory bus of Matrox Parhelia-512 didn't help to overcome the rivals that effectively use their 128bit memory buses.

Nothing new in this test, too. Matrox Parhelia fell behind both: NVIDIA GeForce4 Ti4200 and ATI RADEON 8500.

Faster pixel shaders processing didn't help Matrox Parhelia in this test. Although, thanks to the 256bit memory bus, its performance loss in higher resolutions is smaller than that of NVIDIA GeForce4 Ti4200 and ATI RADEON 8500.

The results of this test suggest that we deal with unfinished drivers. Matrox Parhelia does about two times slower than its competitors and the results remain about the same in any resolution.

In the "Quality" mode of Serious Sam: The Second Encounter the graphics cards deal with more detailed textures and models as well as anisotropic filtering. The results of Matrox Parhelia that can be limited by anything except the GPU performance, drop slightly while NVIDIA GeForce4 Ti4200 evidently didn't like the enabled anisotropic filtering.

The results in Quake3 Arena are mainly limited by the texturing speed and memory bus bandwidth.

In this test, Parhelia-512 couldn't use the advantage of its four texturing units per pipeline again as the test was run with enabled tri-linear filtering. In this mode, the texturing units of Parhelia get combined in pairs and the maximum number of textures rendered per clock drops down to two. The rivals of Matrox Parhelia can also render two textures per clock at most.

NVIDIA GeForce4 Ti4200 and ATI RADEON 8500 feature architectural solutions that increase the efficiency of the available memory bus bandwidth, which allowed them to surpass Parhelia-512 with its broad but ineffectively used 256bit memory bus.

This test uses pixel and vertex shaders and loads the graphics cards quite heavily. Here Matrox Parhelia gains its small revenge thanks to the 256bit memory bus and high pixel and vertex shaders processing speed.

Well, although the Matrox' card could reach the level of ATI RADEON 8500 128MB, NVIDIA GeForce4 Ti4200 is still ahead.

The insufficient CPU performance influences the results of this test a lot: all the cards show simply unacceptable gaming performance here.

However, even under this grave limiting impact of the CPU, Matrox Parhelia fell far behind NVIDIA GeForce4 Ti4200 and ATI RADEON 8500.

The new Matrox chip performed rather poorly in Unreal Tournament Demo, too. Although, in higher resolutions its performance dropped slower than that of the rivals, so that our hero managed to catch up with ATI RADEON 8500.

Anti-Aliasing and Anisotropic Filtering: Quality and Speed

Full-screen Anti-Aliasing and anisotropic texture filtering are functions that don't need the application to involve any additional computational resources. All the extra work is performed by the GPU, so you can increase the image quality in games without loading the CPU that much.

Let's see what Matrox Parhelia has got here.

Anisotropic Texture Filtering

Matrox Parhelia supports anisotropic filtering with the maximum level of anisotropy set to 2. It provides much lower texture precision compared to NVIDIA GeForce4 Ti4600 and ATI RADEON 8500 that support up to 8 and 16 anisotropy level, respectively. Nevertheless, the advantage of anisotropic filtering over bi-linear or tri-linear filtering in Matrox Parhelia can be seen with the naked eye :).

Here's a scene from Quake3 Arena:

The scene from Quake3 Arena with different texture filtering techniques applied. The fragments of the screenshot are on the left, on the right are the same fragments with colored MIP-levels:

Bi-linear filtering		Tri-linear filtering
Anisotropic filtering		Anisotropic + tri-linear filtering

On the screenshots with anisotropic filtering we see that forcing anisotropic filtering also activates tri-linear filtering regardless of what filtering type is set by the application. Here tri-linear filtering was turned on as soon as we forced anisotropic filtering, although the settings indicate that we had bi-linear texture filtering set in Quake3 Arena.

This means that forced anisotropic filtering makes Matrox Parhelia work in the most unfavorable mode. All four texturing units are processing a single texture. Two of them take out two bi-linear samples from one MIP-level of the texture, and the other two take out samples from another MIP-level.

Let's evaluate the results and performance losses in case of forced anisotropic texture filtering in Quake3 Arena:

Now the same in Unreal Tournament 2003 Demo:

The use of forced anisotropic filtering by Matrox Parhelia leads to smaller performance drop compared to NVIDIA GeForce4 Ti4200, but higher performance drop compared to ATI RADEON 8500. By the way, ATI RADEON 8500 cannot force anisotropic filtering in Direct3D with Catalyst 2.03 driver. The performance tests as well as the picture in games indicate it.

So, we could be really happy about the low performance drop by Matrox Parhelia in case of enabled anisotropic filtering. But in these tests all graphics cards worked at different maximum anisotropy level, so this comparison makes sense only if we want to evaluate the performance drop at maximum texture filtering quality. If we set the same maximum anisotropy level for all graphics cards, the picture will be completely different:

Now everything's in its right place. Anisotropic filtering provokes about the same performance drop by both Matrox Parhelia and NVIDIA GeForce4 Ti4200. ATI RADEON 8500 chip uses a very fast anisotropic filtering algorithm, a kind of RIP-mapping, and quite naturally is ahead of all.

Full-Screen Anti-Aliasing

Parhelia-512 chip supports the ordinary 2x2 supersampling as well as a new patented Anti-Aliasing algorithm, FAA-16x, which has been discussed in the first part of the review.

Let's check supersampling first.

As is known, supersampling means processing an increased number of subpixels for pixels located in any part of the screen for the sake of higher texture filtering quality. It works like anisotropic filtering inside the polygons. Let's consider a scene from Serious Sam: The Second Encounter with various combinations of supersampling, tri-linear and anisotropic filtering:

Tri-linear filtering		Anisotropic filtering
Tri-linear filtering + supersampling		Anisotropic filtering + supersampling

Well, supersampling by Matrox Parhelia didn't reveal anything unexpected.

We're more interested in the FAA-16x Anti-Aliasing method. Matrox claims it can provide unrivaled quality of the "jaggies" smoothing at the edges of polygons without significant performance loss.

The results of the tests both confirm and deny the Matrox' words. But, let's keep to the schedule. Let's watch the differences in a scene from WarCraft III with activated supersampling and FAA-16x:

Matrox Parhelia: WarCraft III with FSAA 4x
Matrox Parhelia: WarCraft III with FAA-16x

So, the screenshots suggest the following:

• With FAA-16x enabled, Matrox Parhelia doesn't process all the edges that evidently require the removal of "jaggies".
• The quality of "jaggies" smoothing on the edges where FAA-16x does work is really much better than 2x2 supersampling. The staff of the flag in the first fragment is an example. The inclination of polygon edges is "inconvenient" for 2x2 supersampling and the quality of "jaggies" smoothing is poor here. But with FAA-16x, the "jaggies" disappear completely and are hard to discern even in the enlarged fragment.

We continue. One more scene from WarCraft III:

Matrox Parhelia: WarCraft III with FSAA 4x
Matrox Parhelia: WarCraft III with FAA-16x

Here we see that FAA-16x, unlike supersampling, doesn't make the interface elements "blurred" and that's right.

But in the fourth and the fifth fragments we can clearly see artifacts at the edges of polygons forming the landscape. Seems like we deal with the "loss" of some pixel fragments, that is, FAA-16x works downright incorrectly on some edges.

One more "wonderful" feature of FAA-16x is that it doesn't process the edges of polygons when transparent textures are applies to them. Here's a collection of screenshots from Quake3 Arena that confirm this unpleasant quality of FAA-16x:

Matrox Parhelia: FAA-16x in Quake3 Arena

The screenshots also suggest that FAA-16x works incorrectly with fog, displaying a wireframe model of the skeleton that's peacefully sitting in the corner.

If you're not tired of the enumeration of all the "delights" of FAA-16x, just look at one more screenshot. It's a scene from Serious Sam: The Second Encounter and two fragments of this scene. One - with 2x2 supersampling, the other - with FAA-16x:

Serious Sam: The Second Encounter
Matrox Parhelia: FAA-16x	Matrox Parhelia: FSAA 4x

The screenshots serve to confirm one more evident conclusion: as FAA-16x only processes pixels that lie on the polygon edges, the lines of intersections of polygons wont' be anti-aliased.

In the scene above we see that the lines where the logs touch the water remain "jagged" despite FAA-16x and get smoothed once FSAA 4x is applied.

So, FAA-16x does smooth the "jaggies" on the polygon edges excellently.

But only where it does work. As we've seen not all of the edges that clearly require smoothing are indeed smoothed. Moreover, FAA-16x doesn't work at all at the lines of intersection of polygons and with transparent textures. Moreover, the fog effects in games lead to artifacts and make models wireframe. In some cases, there appear artifacts that look like "loss" or incorrect combination of the processed pixel fragments.

On the whole, I've got an ambiguous feeling about Fragment Anti-Aliasing implemented in Matrox Parhelia. Yes, it's really very high-quality, but provokes too many criticisms on the way. We're left to hope that future versions of the drivers will be able to improve the situation.

Let's see the performance drop once we force FAA-16x and FSAA 4x in Quake3 Arena:

Like supersampling should do, FSAA 4x "kills" Matrox Parhelia's performance, but in higher resolutions, the performance drop is not so big, as in case of enabled multisampling by NVIDIA GeForce4 Ti4200. Seems like 256bit memory bus compensates for it.

Forced FAA-16x leads to lowest performance loss and here Matrox' claims are fully confirmed: Fragment Anti-Aliasing is the most "intellectual" of all the Anti-Aliasing methods available and is accompanied by the lowest performance drop.

The second example is from Unreal Tournament 2003 Demo:

We see the same picture: FSAA 4x by Matrox Parhelia provides absolutely unacceptable gaming performance while FAA-16x makes Matrox the leader.

The last thing to check is the performance of the new card from Matrox and its rivals with forced both anisotropic filtering and Anti-Aliasing. I didn't turn on FSAA 4x by Matrox Parhelia on purpose, because we all know very well what the outcome will be in this case. So, the card was tested with forced FAA-16x:

Very low performance loss in case of activated FAA-16x conceals rather average anisotropic filtering of Matrox Parhelia and allows it to become the No.1.