Professional Affair: OpenGL Graphics Cards Compete – THG.RU

Introduction

Today’s mainstream graphics card market is dominated by products containing Nvidia’s GeForce chip, the TNT2 Ultra, the G400 series from Matrox and ATI’s Rage 128 cards. In our last 3D graphics test we left out one topic: What is actually going on in the OpenGL workstation area?

Professional graphics artists or engineers use hardly any of the graphics chips mentioned above. And this situation is not really improving because of poor support, insufficient driver support for high-end applications and missing functionality. One reason for the poor support are the steadily dwindling profit margins in the last three years. High volumes are the only way to profitability for manufacturers in the mainstream segment. Long gone are the times when workstation PCs sold for 50 000 US-$. This price erosion did not stop at professional OpenGL cards. Currently the ‘sweet spot’ is just below 1000 Dollars. Granted, there are still very expensive OpenGL computing monsters out there, but the number of shipped products over 1000 Dollars is relatively small.

Today’s ‘sweet spot customers’ demand decent performance, stability for many applications, and of course better support than it is usually the case for mainstream cards. In the following comparison we tested five OpenGL cards for less than 1000 Dollars. Initially we were surprised when only the Diamond Fire GL 1 Pro, Evans & Sutherland Lightning 1200, Elsa Gloria II and 3Dlabs Oxygen GVX1 fitted in this price category we selected. 3Dlabs asked us to also test the GVX210 that at almost 2000 Dollars is far above our specification. However, because of its interesting features we decided to include it in our test nevertheless.

Driver direction for OpenGL cards

One thing right at the beginning: If you think that buying an expensive OpenGL graphics card gives you a better performance for games, think again. Generally the graphics cards we tested are optimized for professional applications. Quake players still get a very good price/performance ratio with GeForce graphics cards and should not expect a significantly better performance from our test candidates. On top of that the manufacturers of the products tested below have very little interest in supporting Direct3D games. Accordingly there are not even any Direct 3D drivers for Windows 95/98 and 2000 available for the Diamond Fire GL 1 Pro and the Evans & Sutherland Lightning 1200. The main reason is the OpenGL API for applications and also the operating system architecture. Windows 95/98 does not support multiprocessor systems; for this you need NT 4.0 or the recently launched Windows 2000. Additionally many high-end applications do not run under Windows 95/98. However, even though it might not seem to make any sense from a theoretical point of view, there are quite a few practical reasons for the lack of support for Windows 9x.

1. Windows NT 4 and now Windows 2000 automatically offer a higher system stability, that reduces the support costs for graphic cards manufacturers to a bearable level.

2. Many midsize and large companies changed to Windows NT in the last two years.

3. NT 4 or Windows 2000 (formerly known under NT 5) allow the effective use of the so-called multi-threading. If there is another CPU present the computing processes can be distributed among the processors, increasing the overall performance.

The User Side

We interviewed a few companies that use professional OpenGL applications. We discovered a certain trend during the conversations: Contrary to 3D games, graphics projects hardly have a need for rendering large textures. Modeling of objects often only requires high polygon rates. As long as objects and scenes are still ‘under construction’, many designers are satisfied with displaying it in a wireframe model. And simple shading models are sufficient for viewing individual scenes during the production phase. A higher rendering performance is only necessary when the project enters the last stage. However, we were somewhat surprised when we heard that quite a few 3D scenes in the finishing phase are only rendered by the CPU and not by the graphics card. Many customers demand finished 3D animations in form of a movie presentation. The situation is different though, if the 3D scenes need to be interactive. These simulations follow the same rules as games: high fill rates combined with details requiring high polygon rates are a must. Because the user base is much wider in the OpenGL space than in the classic game environment, the cards are tested with a variety of application and synthetic benchmarks.

Geometry Engine and Rasterizer

If we look at the Evans & Sutherland Lightning 1200 and the two 3DLabs cards, we immediately notice the architectural differences to mainstream 3D cards. The three graphics cards use several graphics chips on the same board. In these products the graphics processors execute the stages of the OpenGL graphics pipeline separately. There also is a differentiation between the geometry engine and the rasterizer. Most of you are probably only familiar with the DirectX terminology from the 3D game world. Therefore we are going to explain the most important terms at this point.

Despite the different terminology the similarity between the Direct3D and the OpenGL pipeline is quite astounding. From a historical point of view the OpenGL API is older than the Direct3D API that is often used in games. But the Direct3D game API, formerly seen as a cheap solution, has come of age. Now it almost offers the same functions that OpenGL has had for years. Nevertheless, Direct3D was unable to conquer the workstation market despite the competitive functions. The main problem: Direct3D does not enable 3D hardware acceleration under NT 4.0. OpenGL, on the other hand, works on all Windows, Unix and Linux platforms. No wonder that programmers of workstation applications support the real standard OpenGL instead of Microsoft’s proprietary Direct3D interface. OpenGL also simplifies porting applications to other platforms.

Both interfaces divide the necessary computing steps into two main areas: geometry and rasterization. The geometry stage converts the coordinate systems for a 3D scene, in relation to the position of the 3D objects in the room as well as the viewing angle (or ‘camera position’) of the user. If necessary, the effects of the light sources must also be calculated. What is simply called geometry in the OpenGL world was named transform and lighting under DirectX and Direct3D, respectively. There are two possibilities for geometry computations under OpenGL: Either the CPU of the workstation runs the calculations, or the computing-intensive steps are outsourced to a special graphics chip, the geometry engine. For quite some time now the OpenGL card manufacturers rely on these geometry engines to take load of the workstation CPU. In the game world under Direct3D this option has been available since the introduction of DirectX 7. Before that the CPU always had to do the task. Formerly nobody was very interested in geometry chips for mainstream cards because of the associated higher costs. This situation changed last year. With the GeForce Nvidia presented the first 3D game chip that integrates geometry and rasterizer on one chip. Shortly after a version for OpenGL cards followed: the Quadro. The latter is used by the company Elsa for the Gloria II.

AGP 2x, 4x and the Pro-Slot

For a few months now there has been an extension for the AGP slot in the true sense of the word. The suffix “Pro” indicates the extension. Both our test motherboards OR840 from Intel and the Asus K7V already have this extension. “Pro” stands for an additional power supply for AGP graphic cards and has nothing to do with the bandwidth of the bus. This extension became necessary because the number of transistors on high-end graphics chips has increased rapidly in the last years and with it also the power requirements. In mainstream graphics cards the standard AGP slot has been sufficient for supplying power to the graphics card up to now. With the 3Dlabs Oxygen GVX210 we saw the first product using the additional voltage supply of the Pro extension.

On the far right, next to the wide gap, are the connections for the additional voltage supply on the 3DLabs GVX210.

Here you see the AGP connection of the smaller brother of GVX210. The GVX1 works without Pro extension.

Can you make out the difference on the Elsa Gloria II? Exactly, the Pro extension is missing here as well, but you see the additional gap in the middle. This enables the transfer mode AGP 4X. All other OpenGL cards in this test are only capable of AGP 2X.

3Dlabs Oxygen GVX1

You can notice the difference to a mainstream game card quite clearly. The GVX1 is filled with computing and memory chips. The eight SGRAM memory modules on the right side of the picture only add up to 32 MByte, used for the frame buffer, the z-buffer and for textures.

The lower part of the picture shows the geometry processor GLINT Gamma G1. It delivers a maximum throughput of 4.75 Million lighted and transformed triangles per second. If necessary, 3DLabs uses its so-called PowerThreads technology. It is based on the OpenGL-ICD, allowing the CPU of the computer to execute a part of the geometry and lighting calculations. In this case the drivers utilize the SSE instructions of the Pentium III or the 3DNow! extensions of the AMD Athlon. The OpenGL drivers are also capable of multiprocessing and load balancing. As you are going to see later in the benchmarks, not every application profits from this technology, but under certain circumstances it is possible to use PowerThreads to optimize the performance. We want to point out here that other manufacturers also programmed load balancing between the CPU and the geometry processor. 3DLabs’ marketing department just cleverly trademarked their method as (PowerThreads(. In principle it is not really that special.

The rasterization processor GLINT R3 is the processor with the largest fan. The R3 includes a 300 MHz Ramdac. It enables ergonomic refresh rates of 100 Hz up to a resolution of 1920×1080 pixel in True Color. The card supports resolutions up to 2048×1536, but then the refresh rate drops to a meager 60 Hz – not something the human eye should suffer for an extended time period. GLINT R3 possesses 7 independent DMA engines and a 128-bit wide memory interface. In our opinion the 32 MByte of video memory is not quite enough. The higher the resolution, the less is available for textures in the video memory of the GVX1. However, the card manages up to 256 MByte external texture memory via the AGP bus in the system memory. Therefore we recommend equipping a motherboard with at least 384 MByte RAM. And here is the first bottleneck already: The Oxygen GVX1 will never be capable of displaying large textures at high speed as soon as the resolution reaches 1600 x 1200. Let’s take a look at the performance at 1280 x 1024.

3Dlabs Oxygen GVX1 – DFP- and Stereo Connectors

Besides the analog standard VGA connector (on the right), there is a DFP connector for digital flat panels. This connector is also known under the name MDR20 and supports appropriate LCD monitors up to a resolution of 1280×1024 @ 60 Hz. If you own a LCD panel with DVI connector, you must buy a DVI-to-DFP adapter available at Foxconn or Molex.

Compared to the maximum (analog) VGA resolution, this DFP resolution is not that high. You should know however, that there are currently no LCD displays with digital connector on the market supporting resolutions higher than 1280×1024. Only NEC offers the 21.1 inch display LCD2110P, that achieves a maximum resolution of 1600×1200. It is only available with an analog VGA connector and not with a digital DFP or DVI connector.

On the far left in the picture you see the 3-pin connector for stereo goggles. Currently Stereographics offers the CrystalEyes goggles. As you can see in the list, up to 60 applications support stereo functionality to display three-dimensional models.

3Dlabs Oxygen GVX1 – Display Characteristics

3DLabs tried to unify the graphics information for its products. We think that the manufacturer went a little too far under Windows NT 4. The section for Direct3D makes little sense because only Windows 9x and Windows 2000 offer hardware support for this API. The software architecture of NT 4.0 does not allow Direct3D acceleration. Otherwise the order of the information is very clear.

The monitor settings are somewhat sparse. Apart from the Gamma controller 3DLabs does not offer anything else.

This tab looks much better. The graphics card driver of the Oxygen GVX1 allows optimizations for different applications like 3D Studio Max, Solidworks, AutoCad or Pro/Engineer – just to name a few. The advanced options are adapted automatically. Tuning freaks may also change the advanced options manually, of course.

Exemplary are the monitor settings. The card automatically reads the data of a DDC compatible display and offers the optimal picture refresh rate. It is also possible to chose between VESA settings and special models with manual settings. 3DLabs is a little too generous with the space. Tabs setup and monitor could have been put together in this case without any problems.

3Dlabs Oxygen GVX1 – PowerThreads

But how does one take advantage of the supposedly great PowerThread capabilities that allow load balancing of the transform & lighting computations between the CPU and the geometry chip on the graphics card?

A 3DLabs employee revealed the secret: The number 16 must be entered in the field “Number of DMA Sub Buffers”. Only then the PowerThread technology fully works, as you can see in the following impressive benchmark figure.

Well, the AWadvs-03 benchmarks show an enormous performance boost. But we could not detect performance increases of this kind in other benchmarks. We were wondering: If the PowerThreads fully unfold with these 16 sub buffers, why is 3DLabs not using them in the preset optimized settings for the applications? If you set the optimization for example to 3D Studio Max, we only read the number 5 in the field. Strange, isn’t it?

3Dlabs Oxygen GVX210

Compared to its little brother the Oxygen GVX210 offers twice as much in many ways. For example this card has two R3 rasterizer chips with twice the fill rates, two VGA outputs and with 64 MByte SGRAM twice as much memory as the GVX1. To say this right at the beginning: we really hesitated to include this card in our comparison test, because it does not meet our specification of a maximum price of 1000 Dollars. 3DLabs sets the price at a hefty 2000 Dollars. The manufacturer shipped the card unsolicited to us. Finally we tested the card nevertheless, since we were interested in the functionality of the second monitor connection, and we wanted to know whether the performance of this card justifies doubling the price.

In the upper section of the picture, you see both the GLINT R3 rasterization processors. As expected, 3DLabs doubles the maximum fill rate from 230 to 460 MTexels/s (dual bilinear mip-map textures) compared to the GVX1. But this does not necessarily mean that it also doubles the performance, as you see later in the benchmark results. The GVX210 utilizes a more powerful geometry processor: the GLINT Gamma 2. It allows a maximum polygon rate of 6.3 MPolys/s (with lighting and transformation). The Gamma G1 on the GVX1 only achieves 4.75 MPolys/s.

Both the blue VGA connectors are clearly visible. Unlike on the GVX1, Oxygen has no space left for the DFP connector. The left side of the picture shows the connector for stereo goggles again, that we already described for the GVX1.

3Dlabs Oxygen GVX210 – Display Characteristics

This part is practically identical to the GVX1.

Again, there are only a few monitor settings available. The Oxygen offers nothing else but Gamma. The only difference to the GVX1 is the tick box for activating the second monitor output. We immediately tried it.

3Dlabs Oxygen GVX210 – Dual-Monitor-Operation

As soon as you click on the small box the tab for the second VGA output appears. As expected it offers the identical features.

The dual operation worked immediately. In this example you see the two views in Solidworks. The left monitor shows the finished rendered model. On the right monitor it can be altered in the design view.

Diamond Fire GL 1 Pro

Since the first version of Diamond(s Fire GL 1 the manufacturer changed owners two times. Diamond, formerly an independent graphics card company, was sold to S3 in fall last year. A few weeks ago S3 disposed of its graphics division by selling it to VIA Technologies. However, the Diamond brand name is supposed to remain despite the change of ownership.

The Fire GL 1 offers very little hardware support for geometry processing, the graphics chip possesses only one rasterizer. Diamond tries to balance this lack of hardware support with clever driver programming.

The new revised IBM graphics chip, that now adds the suffix “Pro” to this graphics card, does not change this situation. However, as you will see in the tests later, Diamond was quite successful in optimizing the driver. The ISV list of tested applications is also impressive.

The Fire GL 1 is equipped with 32 MB SGRAM. The internal Ramdac offers a maximum video bandwidth of 250 MHz. This allows ergonomic picture refresh rates of 85 Hz up to a resolution of 1600 x 1200 pixel. At the maximum resolution of 1920 x 1200 the screen flickers, however, and 75 Hz strains the eyes.

Diamond Fire GL 1 Pro – Display Characteristics

Chip manufacturer IBM was not very creative with the name. It is simply called “256-Bit Graphics Rasterizer”.

Compared to the 3DLabs cards the Gamma and monitor settings of the Fire GL 1 Pro are exemplary. For the basic colors red, green and blue the gamma values can be set separately (the box “Link Sliders” must be deactivated for this). Users, who do not want to navigate through the OSD of their monitor, can simply use the mouse to adjust picture position and size.

The functionality of the picture refresh rates is almost identical with 3DLabs. It is possible to adjust or automatically detect the settings. The latter only works if the monitor supports DDC.

Diamond Fire GL 1 Pro – 3DSM Plugin

Diamond and Elsa offer special driver extensions for 3D Studio Max. Here you see the Fire GL plugin that increases the performance immensely. The 3DSM measurements were done with this plugin.

Evans & Sutherland Lightning 1200

At 399 Dollars the Lightning 1200 from Evans & Sutherland is the cheapest card in this comparison. E & S(s agency was a little hesitant when we approached them. We were told that the Lightning 1200 is not the right choice for our demands – as the price already indicates. The Tornado 3000 would be much more suitable for our test, but the price is 300 Dollars above our specified maximum of 1000 Dollars.

This model really is a little old. The Lightning 1200 still has an external Ramdac chip; on all other cards the Ramdac is integrated in the rasterizer chip. It also only achieves a maximum frequency of 175 MHz. For a comparison: The Quadro chip on the Gloria II delivers a maximum Ramdac frequency of 350 MHz. Consequently the resolutions are somewhat meager: 1280 x 1024 at 85 Hz – that is it. Let’s look at the basic data of this card. The fill rates a far below the average. The maximum texture rate is 70 MPixel/s. Therefore we do not expect too much from this card. To be fair we should therefore test the 1300-Dollar Tornado 3000 at a later point in time. However, it was quite obvious from the beginning that the Lightning 1200 would not do too well. But for CAD users, who only work with simple models, this graphics card is an affordable alternative. On this card E & S uses 15 MByte 3DRAM for the frame & local buffer, 16 MByte CDRAM for textures.

In the bottom half of the picture you clearly see the external Ramdac from TI. The other test candidates offer a Ramdac integrated in the rasterizer. Above are the rasterizer and the programmable geometry unit REALimage 1200.

E & S Lightning 1200 – Display Characteristics

E & S calls the tab with the general information “AccelPanel”. This is the location of the current driver and chip data.

Users who do not understand these settings …

… should use the pre-installed application settings.

Elsa Gloria II with Nvidia’s Quadro-Chip

If you compare this 64 MByte card with the surfboards of the other manufacturers, it seems quite small. However, this impression is deceptive, as you will see in the benchmarks. Elsa does not have its own chip fab but buys them from other manufacturers. Thus the Gloria II uses a customized version of the GeForce graphic chip named Quadro by Nvidia.

While GeForce offers a chip clock frequency of 120 MHz, the Quadro comes with 135 MHz. Therefore the fill rate of the QuadEngine reaches a maximum of 540 MPixel/s instead of the GeForce’s 480 MPixel/s. Consequently the polygon rate also increases from 15 to 17 MPolys/s.

Similar to the GeForce, the Quadro possesses a 350 MHz Ramdac. It allows True-Color resolutions of 2048 x 1536 at ergonomic refresh rates of 85 Hz. A powerful computing monster works under Quadro’s hood. It is the only card in the test that integrates the geometry engine – also know as T&L unit – and the rasterizer on one graphics chip. 3DLabs on the other hand uses a multi-chip solution to achieve the same effect. The card also comes with special drivers for 3D Studio Max and Autocad that were developed by Elsa. The standard drivers are from Elsa as well. The latter gave us a few headaches. More about that soon.

Elsa Gloria II – Display Characteristics

Elsa offers a lot of display extensions. We captured this screen shot from the general info tab.

We do not quite know what this tab is for. Next to it is the standard Windows tab “Settings” that basically offers almost the same functions as the tab “Elsa Settings”.

The color settings are extremely user-friendly. Apart from the RGB controllers for the Gamma, there also is a brightness and contract slider.

This tab contains a variety of preinstalled optimization settings. We will discuss in a moment whether they really offer an advantage.

Elsa and the Driver Problems

If you look for reference drivers for the Quadro chip on Nvidia’s homepage, your search will be in vain. There you only find drivers for the Riva, TNT and GeForce chips. Nvidia refuses any support and points directly to Elsa for Quadro drivers. However, if you look at the INF entries of the GeForce drivers, you discover that GeForce drivers also work with the Quadro. This really is no surprise because the GeForce and Quadro chip cores are almost identical. But before we compare the Elsa Gloria II with all the other cards, we wanted to know whether the Nvidia reference drivers are better than the Elsa drivers. This caused us a lot of headaches, but the good news first: We started the test with the dedicated Elsa MAXTREME driver for the 3D Studio.

Hats off to the engineers at Elsa. They did very good work. In a series of tests the MAXTREME driver more than doubled the performance. The reference driver from Nvidia can only keep up in the Texture3.Max test. In this figure you also see the results from the GeForce chips.

Elsa Drivers – The Bad News

If you use other OpenGL programs besides 3D Studio Max or Autocad, you must go back to the Elsa standard drivers. The dedicated plugins do not help anymore. Of course you would expect drivers that achieve maximum performance. We now use the synthetic benchmark Viewperf 6.1.1 to compare two of the latest Elsa drivers 102 and 116 on the Asus K7V with Nvidia’s Detonator drivers 5.13.

While the individual tests with the Nvidia drivers run through very fast, we were twiddling our thumbs with the Elsa drivers. Do you notice the massive performance drops in almost all categories? Even a GeForce with Detonator drivers shows a better performance than the Quadro on the Gloria II. The new 116 drivers are even slower in Viewperf than the older 102 drivers.

This trend continues in a real application as well. In Solidworks 99 the Nvidia reference driver leaves the Elsa drivers in the dust. At least the 116 Elsa driver is a little bit faster than its predecessor 102.

In the Fogcity benchmark the Elsa drivers again displayed the weakest performance. We discovered another interesting point: Even though the GeForce has less memory and a slower clock rate, it is able to profit from the higher bandwidth of the DDR memory. The GeForce-DDR card is more than 50 percent faster than the Gloria II with Elsa drivers.

Quality Problems

And another point made us feel uncomfortable: Elsa’s 4.02.04.102 drivers are faster in Viewperf than the succeeding 116 drivers, but there are quality problems.

In both screenshots you notice the strange wireframe models that ‘shine’ through the scenes in Solidworks 99 rendered with Elsa’s two 102 drivers.

This is how the scenes are supposed to look. The 116 finally eliminates the problem. The overall performance of the Elsa drivers is still poor if compared to the Nvidia reference drivers. Even when compared to each other the Elsa drivers perform inconsistently. In Viewperf the 116 is slower than the 102.

Table of Features

The Test Platforms

We tested all cards on two 800 MHz platforms. One was based on Intel’s workstation board OR840 that we equipped with 384 MByte PC800 RDRAM. The processor was a Coppermine Pentium III 800EB. The Intel chipset i840 supports two Rambus channels, allowing doubling the maximum bandwidth to 2 x 1.6 GByte/s.

Alternatively we tested on the much cheaper Athlon 800 platform with 384 MByte PC133 SDRAM. We used the Asus K7V with KX133 chipset from VIA as motherboard. The maximum SDRAM memory bandwidth is 1.1 GByte/s.

Test Procedure

We used synthetic and application benchmarks for the performance evaluation of the cards. For the monitor resolution and color depth we followed the recommendations of the Standard Performance Evaluation Corporation, short SPEC.

We tested all cards under Windows NT 4.0, Service Pack 6a with a resolution of 1280×1024 and True Color. Exemption: Because of the driver problems with the Gloria II, we ran 3D Studio Max with the special Elsa MAXTREME driver; all other tests were done with Nvidia’s Detonator reference driver 5.13.

Synthetic Benchmarks

For the synthetic benchmarks we used SPECopc SPECviewperf 6.1.1, Glaze and Fogcity.

Application Benchmarks

Real world benchmarks that run real software programs are an important criterion. We chose the SPECapc Solidworks 99 Benchmark and the R3-Benchmark-Suite for 3D Studio Max 3.1.

3D Studio Max 3.1 R3 Benchmark

In this test the camera is constantly moving through the 3D scene. The graphics card must recalculate the scene completely for every frame. In relation to the resolution this extreme rasterization test works with large and small polygons simultaneously.

The clear winner in the rasterization category is the Fire GL 1 from Diamond, achieving an average of 57 frames per second. The Elsa Gloria II with the Quadro chip from Nvidia finishes in second place. All results on the Athlon 800 platform are 5 frames/s below the results on the OR840 with the Pentium III 800. As expected, doubling the fill rate with two GLINT R3 rasterizer chips on the Oxygen GVX210 does lead to twice the performance if compared to the regular fill rate of the GVX1. But at least by implementing the GLINT R3 chips 3DLabs achieves a performance increase of about 45 percent over the smaller brother GVX1.

Geom2.Max uses scenes with up to 200000 polygons. This resembles a good mixture of large connected and unconnected objects. It is an excellent test for the geometry features of the graphics cards. Elsa’s Gloria II takes the lead, with the especially programmed MAXTREME plugin boosting the performance of 3D Studio Max. The Nvidia detonator drivers would have strangled the performance of the Gloria II mercilessly, allowing 3DLabs to capture the first and the second place. Nevertheless, we think that the drivers are programmed sloppily. On the Athlon platform the Gloria II, GVX1 and GVX210 achieve only insufficient frame rates. The Fire GL 1 Pro and E & S Lightning 1200 are more balanced. Even though the overall performance is not as good, the performance gap between the Pentium III and the Athlon platform is smaller percentage-wise.

3D Studio Max 3.1 R3 Benchmark, Continued

This benchmark tests the accelerated lighting capabilities of the graphics card. It uses 8 Omni light sources.

We discovered an interesting effect here. Theoretically the Quadro should have won the test. After all it is equipped with a powerful Transform & Lighting unit. But the Fire GL 1 Pro tricks the chip. This card does not even have any hardware acceleration for light sources, requiring the CPU to do the calculations. But with intelligent driver programming Diamond takes advantage of the SSE of the Pentium III and 3DNow! of the Athlon, respectively, and outmaneuvers the Gloria II.

In this benchmark 3D Studio Max displays four viewports simultaneously. The objects are visible from four camera positions at the same time. Again, Elsa’s MAXTREME driver does the best job and leaves all other test candidates far behind. Especially the Lightning 1200 and Fire GL 1 show a very poor rendering performance is this scenario.

3D Studio Max 3.1 R3 Benchmark, Continued

In Texture3.MAX the graphics card must calculate dynamically changing geometry with a huge texture map. It is almost spooky, but in case of 3DLabs and Nvidia the texture fill rates specified by the manufacturer follow the frame rates almost linearly. As expected the Elsa Gloria II with its Nvidia Quadro chip takes the lead, closely followed by the Oxygen GVX210.

This test uses complex wireframe models without shading or textures. CAD users who often create technical drawings will find this to be a meaningful way of comparing the features of the test candidates. After the last benchmark figure you are already aware of the first trend. Because of its MAXTREME plugins the Elsa Gloria II is in the lead in almost all categories. The Quadro chip stands out against the other models distinctly. In relation to that the differences for displaying a pure wireframe model are not as significant among the other four graphics cards.

3DSM – the other results

We decided to only publish the most representative tests from 3D Studio Max in a diagram. All other results are displayed in this numeric table.

SPECapc Solidworks 99 Benchmark

The second application benchmark suite is from SPEC. This organization also offers the synthetic benchmark Viewperf. All of the following results are not frame rates but refer to a reference machine equipped with a Pentium II 300 and the Permedia 2 graphics processor.

The Solidworks 99 benchmark includes different tests for CPU, I/O and graphics subsystem. We only used the three graphics tests. The scenes have up to 276000 polygons. According to the SPECap specifications the models reflect the typical construction work of a Solidworks user. The benchmarks hardly contain any textures. Instead it generates complex production facilities like an assembly line with simple shadings (for example Gouraud Shading) and lighting effects.

Bottom Case uses relatively simple scenes without any graphical frills. The Fire GL 1 Pro achieves a relatively high scoring of 3.8 on the Pentium III platform.

As soon as the models get more complex the performance of the Fire GL 1 Pro collapses. The Quadro on the Gloria II flexes its muscles again – but it hardly can loose the GVX210 or the GVX1. 3DLabs needs to update its drivers. The results on the Athlon platform could be better.

If the transparency effects are switched on, Athlon’s capabilities improve a little. Nevertheless the order remains the same as in the previous test: Gloria II, Oxygen GVX210 and GVX1.

SPECopc SPECviewperf 6.1.1

On the i840 platform with the Pentium III 800 the Gloria II achieves impressive 91.5 frames per second, and reaches the fill rate limit of the 135 MHz Quadro chip. On the KX133 platform with Athlon 800 the Quadro is still in first place, but the performance drop is unusually high. The NT 4 drivers seem to be the cause, because under Windows 98 the delta between Athlon and Pentium III for the Advanced Visualizer is only minimal.

Again the KX133 shows a weaker performance than the RDRAM platform from Intel. But the graphics card drivers also need improvements – apart from the ones for the Fire GL 1 Pro. Fortunately the differences are less apparent in the application benchmarks.

SPECopc SPECviewperf 6.1.1, Continued

The Data Explorer Text DX-05 repeats the same poor results for the graphic cards on the KX133 Athlon platform as the DRV-06 (Design Review).

The ray-tracing and radiosity computations from Lightscape require a lot of floating point performance. Because of its superior FPU the Athlon processor could help out a lot here. The Fire GL 1 Pro is the only card in the test that proves this theory. Otherwise we again criticize the poor driver capabilities of the other test candidates regarding the KX133/Athlon.

The extreme lead of the Fire GL 1 Pro is due to the extensive use of the Streaming SIMD extensions of the Pentium III. Intel should appreciate this very much. The Gloria II and Fire GL 1 show almost identical results. We wonder whether the graphics cards plays a role at all. Does the Pentium III do all the computations?

Fogcity

The Fogcity benchmark is packed with polygons and bandwidth consuming fog effects. These types of simulations are the strength of the Oxygen GVX210. Just remember the earlier benchmarks: If the Gloria II was equipped with DDR memory, it could have easily won this test.

Glaze

Glaze was developed under the patronage of Evans & Sutherland. It consists of 15 individual tests that we used to determine the geometric average for the frame rate. If you compare these results with all other benchmarks, cards like the Oxygen GVX1 do not follow the general trends. At the top are the GVX210 and the Gloria II again. We do not use the values for an overall evaluation because of a few inconsistencies.

Recommendations and Resume

Generally the Gloria II from Elsa has the best potential to win in this price category. However, we presently cannot recommend this Quadro graphics card. The main problem is the driver situation: The Elsa standard drivers are very stable but the performance is unsatisfactory. Using Nvidia Detonator drivers instead reveals the actual power of this card. Unfortunately there is no official support for the Nvidia drivers and another drawback are the instabilities we observed on the Athlon and the Pentium III platform. If you only work with Autocad or 3DStudioMax, you can buy this card right now. Elsa offers dedicated drivers for these programs that guarantee maximum performance and stability.

For a creator and typical CAD/CAM user the Fire GL 1 Pro is a very finished product. But do not expect that you can display complex simulations or animations. When trying the latter the performance would surely collapse, because the Diamond card neither has a geometry chip nor a powerful texturing unit.

The Oxygen GVX210 from 3DLabs offers you the choice of connecting a second monitor. This card is suited for almost all high-end applications on Pentium III platforms, it even comes with the option of purchasing stereo goggles. 3DLabs should fine-tune the drivers for the Athlon platforms, because only on the Intel OR840 motherboard the performance was up to par. The only stumbling block is the price. At 2000 Dollars the card is too expensive for the performance it delivers.

Like the Fire GL 1 Pro we recommend the quite affordable Oxygen GVX1 only to CAD/CAM users. Contrary to the Diamond card the GVX1 possesses a geometry chip. As long as the number of polygons is not too high, the card even handles simple simulations.

We do not really know what to make off the E & S Lightning 1200. It ends up in last place in almost every benchmark. Evans & Sutherland seems to be aware of the performance problem of this older card, and has lowered the price. This model only costs 399 Dollars. At a later point in time we should take a look at the Tornado 3000 that supposedly offers a much higher performance. With 1300 Dollars it is above our specified price point, however.

Pros and Cons

The following table shows a summary of the advantages and disadvantages of the tested cards.

Final Thoughts: Rambus-Pentium versus SDRAM-Athlon

This test clearly reveals better performance values for RDRAM platforms with the Pentium III. Contrary to the game PC world, the workstation segment finally gives Rambus memory the chance to show off its performance advantages propagated by Intel. In comparison to the Athlon 800 we measured better results in almost all categories at the same clock frequency under Windows NT 4.0.

Some card manufacturers seem to prefer the Pentium III for the driver development. In the application benchmarks the difference is less noticeable than in the synthetic benchmarks. We generally assign more importance to the application benchmarks because they reflect the real-life situation for the user much better. On top of that Nvidia proves that it is possible to optimize graphics card drivers for the Pentium III and Athlon equally.

One point is quite important when deciding on a platform. Workstations are generally equipped with more memory than mainstream PCs. 512 MByte are no rarity. Given today’s prices, RDRAM would be much more expensive than the rest of the hardware. In this case the Rambus memory alone adds up to a whopping 5000 Dollars, five times more than PC133 SDRAM. The slight performance disadvantages of the Athlon KX133 platform can be balanced with more MHz. This would be a much wiser investment for your money.