Introduction
Recently it seems as if NVIDIA has become pretty much omnipresent and almost omnipotent in each of the different graphics sectors, from the low-end/ high-end desktop to the workstation and mobile arena. However, NVIDIA has not managed to change the basic rules of the professional OpenGL-graphics market yet. Here, NVIDIA’s short product cycles are neither necessary nor particularly welcome, because workstations are company investments that aren’t supposed to get upgraded every half-year. Nobody really cares about fancy pixel or texel fill rates or features like FSAA and DX8-support either. The price of a professional OpenGL-card is not of major importance, because in the CAD-field the actual engineer is what is really costly and so those companies are willing to pay big money for whatever makes the work for their engineers easier and quicker.
Diamond’s German-based professional OpenGL-group had already been one of the major developers of top-notch OpenGL-accelerators long before NVIDIA made its entry in the professional OpenGL-market with ‘Quadro’, a chip identical to GeForce, but running with a few different hardware switches and different drivers, which received its high performance from NVIDIA’s by-the-time brand new integrated T&L-engine. The last chip in Diamond’s long FireGL-line of products was the FireGL1-card, which had finally been beaten by NVIDIA’s Quadro. Before the release of NVIDIA’s first OpenGL-accelerator the FireGL1 had pretty much ruled the segment of midrange-priced professional OpenGL-cards, exclusively equipped with a chip designed and made by IBM. There’s a cloak of secrecy and mysticism over this OpenGL-chip development group of IBM, which had remained pretty quiet until recently, when it finally launched a new line of chips that is again exclusively found on Diamond’s new FireGL2-card. This new card seems well potent to challenge NVIDIA’s latest professional OpenGL-solution by the name of ‘Quadro 2 Pro’, which is exclusively found on the Gloria III card from Elsa. Either solution comes at a retail price of about 1200 dollars.
The R&D teams of NVIDIA and Diamond’s OpenGL-group are adherers of two completely different philosophies. NVIDIA, the uncontested leader in high-end graphics chips for the 3D desktop/gaming market, is using its latest chip variety to establish itself in the OpenGL workstation market. The Quadro2 Pro chip is mainly based on the GeForce2 GTS core but runs at higher clock speeds, uses faster memory and offers a few more (again hardware switched and software enabled) features.
Former exclusive partner ELSA decided to unite its graphic R&D team resources with NVIDIA. This means ELSA does not develop graphics cards by itself anymore but uses NVIDIA’s reference designs instead. Elsa’s and NVIDIA’s workstation driver developers are cooperating as well. The Elsa’s Gloria III card is basically the NVIDIA-designed Quadro2 Pro reference card.
The designers from Diamond’s Starnberg/Bavaria-based professional OpenGL-group come from a different background. They are primarily focused on the OpenGL workstation market without paying any attention to 3D-gaming.
In this test we cared little about the politics of the abovementioned companies, but were merely interested in their current products. Thus, we have the FireGL2 compete with NVIDIA’s Quadro2 Pro reference card (aka Elsa Gloria III). As we’ve already published test results of the Quadro2 Pro in previous articles (New NVIDIA Detonator 3 Drivers For Linux, NVIDIA 3D Under Linux) and also use this card for the SpecViewperf benchmarks, we decided to focus on the features of Diamond’s FireGL2.
FireGL2 Features
Large coolers and no fans – that was our first impression of the FireGL2. Diamond offers a simple explanation for this. Most of its OEM customers do not want a fan on their graphics card. Thus, the coolers need to be designed with more surface area in order to dissipate heat effectively enough. Regrettably, this means that there’s hardly any headroom left for overclocking unless you equip this card with an HSF-combination.
Here we see three connectors. From left to right, there are: Stereo MDIN3 for 3D glasses, DVI-D for digital flat panel displays and D-Sub VGA for analog CRT displays.
The three-pin connector on the left conforms to the specifications of the CrystalEyes 3D glasses from Stereographics. The DVI socket only provides digital signals, analog signals as with DVI-I are not provided. This means that you can’t connect two analog displays (CRTs) to the GL2. The RAMDAC-bandwidth of the FireGL2 is set to 300 MHz, enabling resolutions up to 1600 x 1200 pixels at 85 Hz refresh rate. Maximum resolution of 1920 x 1200 can only be run at 76 Hz, however.
FireGL2’s RC1000 Rasterizer Chip
FireGL2 is equipped with two graphic chips. The larger of the two coolers covers the RC1000 rasterizer, which is clocked at 120 MHz.
The chip’s die is covered with some kind of copper heat spreader. The remarkable thing about RC1000 is its diagonal mounting to the board.
The RT1000 offers the following features:
- 2D/3D Raster Engine with Texture Unit clocked at 120MHz
- Video Overlay Unit
- 2 DMA / BLT Units
- Polygon Setup Processor
- 300MHz Ramdac (30 bit Palette DAC)
- 4 tables for Color Lookup and Gamma Correction
- 256-bit wide memory interface for DDR memory
Compared to Quadro2 Pro, the theoretical fill rates of the RC1000 are not impressive. The chip manages a meager 410 Mpix/sec. The texture fill rate is even lower, hovering around 200 Mtex/sec. In comparison, Quadro2 Pro with 1000 MPixel/sec and 2000 MTexel/sec is a true Goliath. We’ll discuss the (minor) importance of fill rates in professional OpenGL applications later in this report.
FireGL2’s GT1000 Geometry Chip
The GT1000 chip is responsible for the T&L part of the OpenGL pipeline. It runs at 190 MHz and is thus clocked higher than the RC1000-chip. On the Quadro2 Pro both rasterizer and T&L unit are integrated on one chip. Both engines run at 250MHz on the NVIDIA product. According to Diamond, the geometry engine manages 35 Gflops/sec. This translates into a polygon rate of 27 Mpolys/sec. On paper, the NVIDIA chip seems to have the upper hand here as well, considering it is rated at 31 Mpolys/s. While the Quadro2 Pro shows certain remnants of 3D-game optimizing, Diamond claims to have inserted a “100% OpenGL Geometry Pipeline” into their GT1000.
Feature Overview:
- Full Geometry Transform Processing
- Hardware Calculation of up to 16 light sources (more sources through software calculation)
- Calculating Texture Coordinates and Fog
- Multi Texturing Support
- Hardware Clipping
- Immediate Mode Rendering
- 190 MHz Clock
Who’s Right? Customer Wishes Perceived
The FireGL2 doesn’t care much for Windows 98 or ME. You will search in vain for drivers that would support the two ‘toy-OSes’. This is due to the market position of the card as a pure OpenGL workstation product. Diamond drivers are currently restricted to Windows NT 4.0, Windows 2000 and UNIX – the OSes most widely used for workstations. From the beginning, Diamond offers driver optimizations for the most important command set extensions of current CPUs. 3DNow! is implemented for AMD Athlon, and for Pentium III and 4 we have SSE and SSE2 support respectively.
A look at the FireGL2 specs is disappointing at first, as fill rates are rather low compared to those of the Quadro2 Pro. Still, there’s quite a competitive polygon rate. We interviewed several different companies that use professional OpenGL applications on a daily basis. During talks, we noticed a certain trend: While working on a graphics project, rendering of large textures is much less important than in 3D games, where it has become most essential. However, the modeling of objects requires very high polygon rates. As long as objects and scenes are still “under construction”, most designers are handling them as wire frame models. Even looking at single views or scenes is done with simple shading models during production phase. Only at the final stage of a project the designer requires some kind of high rendering performance and thus fill rate. It shows that the user profile of a construction designer is completely different from that of a 3D-gamer. This explains why Diamond put little effort into the rasterizer. But even lay(wo)men can appreciate the connection between a wire frame model and the modeling process. Many of you might have seen Spielberg’s dinosaur-movies. In “The Making of” you get a look behind the scenes and developers explain how the dinosaur-models were developed. You can spot it right away that the animations were exclusively done using wire frame models. Way towards the end after the scene sequence has been determined, finally the wire frame dinosaurs are ‘dressed up’ with complex textures and included into the movie.
Whether Diamonds philosophy paid off or not can be seen from the benchmarks in the next section.
Display properties
The drivers add only three tabs to the display properties of Windows. Compared to NVIDIA, this looks very “tidy”.
Here you see the current driver version. Memory size is reported as well as clock rates of RAMDAC and memory.
Most important here is the gamma correction tab, which is decisive for graphics designers. Separate adjustment of red, green and blue is possible as well as a combined adjustment via the “Link Sliders” checkbox.
Display Properties, Continued
Optimizing The FireGL2 For Certain Applications
Depending on your preferences, you can optimize FireGL2’s driver for different applications. For completenessґ sake we’ve listed all applications in this overview.
- 3D Studio Max / VIZ 3.x
- 3D Studio Max / VIZ 2.0
- 3D Studio Max / VIZ 2.5
- AutoCAD
- AutoDesk Inventor
- CATIA Version 5
- CDRS / 3DPAINT
- EUKLID
- Helix Design System
- Houdini
- I-DEAS Master Series
- ICEM Surf
- Lightwave 3D
- MAYA NT
- Microstation
- MultiGen Creator
- Pro Engineer
- Quad Buffer Stereo
- Softimage 3D / DS / XSI
- SolidDesigner
- SolidWorks
- TrueSpace
- Unigraphics
Feature List
Manufacturer | Sonicblue/Diamond | NVIDIA/ELSA | NVIDIA |
Graphics Card | FireGL2 | Quadro2 Pro | GeForce2 GTS |
Memory | 64 MB DDR SGRAM | 64 MB DDR SGRAM | 32 MB DDR SGRAM |
AGP interface | AGP 4X | AGP 4X | AGP 4X |
1st display connector | VGA D-Sub (analog) | VGA D-Sub (analog) | VGA D-Sub (analog) |
2nd display connector | DVI-D (digital) | DVI-I (analog & digital), simultaneous operation not possible | optional |
Video Ramdac | 300MHz | 350MHz | 350MHz |
Geometry chip | IBM GT1000 @190MHz | Quadro2 Pro T&L @250MHz | GeForce2 GTS T&L @200MHz |
Rasterizer | IBM RC1000 @120MHz | Quadro2 Pro 3D @250MHz | GeForce2 GTS 3D @200MHz |
Memory clock | 120MHz DDR | 200MHz DDR | 166MHz DDR |
Pixel fillrate | 410Mpix/s | 1000Mpix/s | 800Mpix/s |
Texture fillrate (trilinear filtered) | 200Mtex/s | 2000Mtex/s | 1600Mtex/s |
Polygon rate | 27Mpolys/s | 31Mpolys/s | 25Mpolys/s |
MPEG Motion Comp | no | no | no |
Drivers | |||
Windows 2000 | yes | yes | yes |
Windows NT 4.0 | yes | yes | yes |
Linux | yes | yes | yes |
Windows 98 | no | yes | yes |
Optimizations for additional command sets | |||
Pentium III SSE | yes | yes | yes |
Pentium 4 SSE2 | yes | yes | yes |
Athlon 3DNow! | yes | yes | yes |
Test Platforms
We let a FireGL2 and a Quadro2 Pro reference card (identical to Elsa Gloria III) compete against each other. Additionally, we show test results for a mainstream GeForce2 GTS chip for better assessment of both products. We used three different platforms for each card: one with AMD Athlon 1.2 GHz, Intel Pentium III 1.0 GHz and Pentium 4 1.5GHz. In our benchmark charts you’ll also see how a Pentium 4 would work if it ran at 1.2GHz (same clock as AMD Athlon 1.2GHz).
Graphics Cards Tested | |
Diamond FireGL2 | |
Graphic chips | IBM RC1000 (rasterizer) and IBM GT1000 (geometry) |
Memory | 64 MB DDR SGRAM |
Memory clock | 120MHz DDR (or “240MHz” in marketing language) |
RC1000 clock | 120MHz |
GT1000 clock | 190MHz |
BIOS version | 1.04 |
Driver version | 5.12.2195.2044 |
NVIDIA Quadro2 Pro reference board (identical to Elsa Gloria III) |
|
Graphic chip | Nvidia Quadro2 Pro |
Memory | 64 MB DDR SGRAM |
Memory clock | 200MHz DDR (or “400MHz” in marketing language) |
Chip clock | 250MHz |
BIOS version | 3.15.0006 |
Driver version | Nvidia 647 |
ASUS AGP-V7700 | |
Graphic chip | Nvidia GeForce2 GTS |
Memory | 32 MB DDR SGRAM |
Memory Clock | 166MHz DDR (or “333MHz” in marketing language) |
Chip clock | 200MHz |
BIOS version | 2.15.0111 |
Driver version | Nvidia 647 |
Test Platforms | |
HDD for all plattforms | Quantum Fireball lct08 12 GB |
Athlon (PC2100 DDR SDRAM) | |
Mainboard | Gigabyte GA-7DX Rev.2.3 |
Chipset | AMD760 |
BIOS version | F1a |
CPU | AMD Athlon 1.2GHz at 133MHz Double Pumped System Bus |
Memory | 128MB Micron CL2 DDR SDRAM (1 module) |
Pentium III (PC133 SDRAM) | |
Mainboard | ASUS CUSL2 |
Chipset | Intel 815 |
BIOS version | 1003-3 |
CPU | Intel Pentium III 1.0GHz at 133MHz Single Pumped System Bus |
Memory | 128MB WorkX PC133 SDRAM (1 module) |
Pentium 4 (PC800 RDRAM) | |
Mainboard | Intel D850GB |
Chipset | Intel 850 |
BIOS version | V1 |
CPU | Pentium 4 1.5GHz at 100MHz Quad Pumped System Bus |
Memory | 128MB Samsung PC800 RDRAM (Dual Channel Rambus, 2 modules) |
Software and Settings | |
OS version | Windows 2000 Professional, SP1 |
Solidworks 99 | SPECapc SW99 benchmark |
SPECviewperf | Version 6.1.2 |
3D Studio Max 3.0 | R3 benchmark |
Resolution | 1280×1024@32-bit color depth |
V sync | “off” in all benchmarks |
Test Methodology
We used synthetic and application benchmarks for assessing performance of these cards. Screen resolution and color depth were chosen according to the recommendations of the Standard Performance Evaluation Corporation, SPEC. We tested all cards under Windows 2000 Service Pack 1 with a resolution of 1280×1024 and True Color. Vsync is set to “off”, to avoid its frame rate impact.
As synthetic benchmark we used Viewperf 6.1.2. Real Word benchmarks using real software are another very important criteria for evaluation. We decided to use the SPECapc Solidworks 99 Benchmark and R3-Benchmark-Suite for 3D Studio Max 3.0.
3D Studio Max 3.0 R3 Benchmark
This test has a camera that’s constantly moving within a 3D scene. For each frame the graphics card has to recalculate the entire scene. In sync to the resolution, this extreme rasterization test works with big and small polygons simultaneously.
3D Studio Max 3.0 R3 Benchmark, Continued
The undisputed winner in the Rasterization category is Quadro2 Pro from NVIDIA, which reached a whopping 112 fps on the Athlon 1.2 GHz platform. This is not surprising, as the Quadro2 Pro shows a texture fill rate that is ten times higher than that of its counterpart.
The scenes of Geom2.MAX uses up to 200,000 polygons. They contain a good blend of connected as well as separate objects and are therefore an excellent test of graphics card T&L-capabilities. The FireGL2 by Diamond takes the lead, even though its theoretical polygon fill rate is supposed to be slightly lower than that of the Quadro2 Pro.
3D Studio Max 3.0 R3 Benchmark, Continued
This benchmark tests whether the graphics card can accelerate lighting. 8 ‘omni’ light sources are implemented. Using hardware only, the FireGL2 can calculate up to 16 light sources without the need to fall back to software algorithms. Quadro2 Pro lags behind in this category as well.
In this benchmark 3D Studio Max renders four view ports at once. This means that objects are shown from four different (camera) viewing angles. Quadro2 Pro manages only one third of FireGL2’s performance.
3D Studio Max 3.0 R3 Benchmark, Continued
Texture3.MAX requires the graphics card to use a giant texture map for calculating constantly reshaping geometry. In this test Quadro2 Pro and even GeForce2 GTS are able to show their superior texture performance.
This test uses complex wire frame models without any shading or texture. CAD users who often create technical drawings will find this a very revealing test for pure construction work. With wire frame models, Quadro2 Pro has no chance at all.
3DSM – The Remaining Results
We decided to show a graphic depiction of only the most representative 3D Studio Max tests. All the other test results are presented in this numerical table:
3DstudioMax3.0 Blttest | PIII 1GHz | Athlon 1.2GHz | P4 1.2GHz | P4 1.5GHz |
GeForce2 GTS | 268 | 321 | 244 | 268 |
Quadro2 Pro | 275 | 335 | 250 | 272 |
FireGL2 | 383 | 413 | 357 | 382 |
3DStudioMax3.0 Geom1 | PIII 1GHz | Athlon 1.2GHz | P4 1.2GHz | P4 1.5GHz |
GeForce2 GTS | 9.4 | 11.9 | 10.5 | 11.5 |
Quadro2 Pro | 9.4 | 12.1 | 10.5 | 11.5 |
FireGL2 | 20.6 | 26.8 | 33.1 | 35.9 |
3DStudioMax3.0 Light1 | PIII 1GHz | Athlon 1.2GHz | P4 1.2GHz | P4 1.5GHz |
GeForce2 GTS | 41.0 | 41.7 | 39.8 | 40.3 |
Quadro2 Pro | 49.6 | 50.7 | 45.0 | 45.4 |
FireGL2 | 75.4 | 83.2 | 79.9 | 81.9 |
3DStudioMax3.0 Light2 | PIII 1GHz | Athlon 1.2GHz | P4 1.2GHz | P4 1.5GHz |
GeForce2 GTS | 55.9 | 58.3 | 54.7 | 55.5 |
Quadro2 Pro | 53.1 | 70.4 | 54.9 | 55.5 |
FireGL2 | 75.0 | 83.2 | 79.9 | 81.9 |
3DStudioMax3.0 Texture1 | PIII 1GHz | Athlon 1.2GHz | P4 1.2GHz | P4 1.5GHz |
GeForce2 GTS | 72.5 | 56.7 | 77.0 | 77.6 |
Quadro2 Pro | 111.4 | 115.9 | 80.9 | 82.4 |
FireGL2 | 51.8 | 52.1 | 50.8 | 51.4 |
3DStudioMax3.0 Texture2 | PIII 1GHz | Athlon 1.2GHz | P4 1.2GHz | P4 1.5GHz |
GeForce2 GTS | 47.4 | 50.4 | 48.2 | 48.7 |
Quadro2 Pro | 47.6 | 59.4 | 40.0 | 40.9 |
FireGL2 | 36.8 | 37.1 | 36.4 | 36.7 |
SPECapc Solidworks 99 Benchmark
Results of this application benchmark are not frame rates but refer to a reference system equipped with Pentium II 300 CPU and Permedia 2 graphics engine.
We only used the OpenGL graphics test out of the Solidworks 99 suite. The polygons shown in these scenes can number up to 276,000. According to SPECapc, these models are supposed to represent the typical construction processes undertaken by Solidworks users. Regardless of platform, there’s little difference between FireGL2 and Quadro2 Pro. Still, the latter shows worse results.
SPECopc SPECviewperf 6.1.2
Advanced Visualizer is an exhaustive test for modeling, animation, rendering and image composition. FireGL2 turn out to be about 30% faster than Quadro2 Pro in this synthetic test.
Design Review is a program combination that visualizes three-dimensional elements, such as tubes, brackets and other construction elements used in complex models. Again, FireGL2 takes the lead, shining especially in the Pentium 4 platform.
SPECopc SPECviewperf 6.1.2, Continued
Data Explorer is an IBM product. It encompasses the analysis and visualization of object data. At best, FireGL2 is twice as fast as Quadro2 Pro.
The Ray tracing and radiosity calculations used by Lightscape require a whole lot of FPU-power. Though both OpenGL solutions have powerful graphic chips, the CPU’s floating-point unit is what really matters. Athlon is able to show its muscles.
SPECopc SPECviewperf 6.1.2, Continued
Half of all ProCDRS tests are based upon wire frame models, some of which have up to 262,000 triangles. The remaining subtests still have a strong wire frame component, with simple shading surfaces added to the frame. The CPU is of little importance here, as graphic chips have to carry the major workload. FireGL2 is approximately 17% faster than Quadro2 Pro.
MedMCAD_01 models objects typically appearing in ProEngineer (PTC) and Solidworks. FireGL2 holds only a slight advantage over Quadro2 Pro.
Conclusion
Diamond FireGL2 takes the lead over the Quadro2 Pro card in the majority of benchmark categories. This test proves that the crucial point with OpenGL is not high polygon and fill rates, as with 3D gaming.
The FireGL developer team from Starnberg/Germany did an excellent job adjusting hardware to professional workstation applications and optimizing drivers for AMD and Intel platforms.
Diamond tested an impressive amount of applications as well. The conclusion is therefore rather simple. Diamond’s FireGL2 definitely deserves our recommendation.
The FireGL2 has a good price/performance ratio, and all necessary connectors are in place: We find analog VGA Out, a DVI I connector for digital flat panels and a socket for the CrystalEyes-spectacles from Stereographics. The recommended retail price of US$1200 leaves FireGL2 within the “sweet spot”.
Looking at the Quadro2 Pro reference card aka Elsa’s Gloria III, NVIDIA’s origin from 3D gaming becomes quite obvious. NVIDIA Quadro 2 Pro was able to gain considerably over last year’s Quadro chip, but Diamonds equally priced product is able to show NVIDIA who is boss. The only time FireGL2 demonstrates some weaknesses is with texture-intensive models.