A Gourmet's Review of NVIDIA's GeForce2 Ultra
Oh, my dear readers, there are no words to express how that meal - the glutinous texture of which reminded me of some bottled mountain lizard - I had the ill fortune to taste in a picturesquely half-timbered country inn in a distant region lay heavy on my stomach. But no sooner had I recovered from this misery than - as a compensation for my sufferings - my Editor in Chief granted me a commission I had been looking forward to for ages. This time I had to visit a really distinguished posh restaurant: The NVIDIA! I knew it quite well from the enthusiastic reports of my colleagues that this place has some real specialties. Needless to say, I was especially interested in the latest tidbit offered by the restaurant: the GeForce2 Ultra á la NVIDIA. So I have determined beforehand that I am going to choose from the menu a GeForce2 MX as an hors-d oeuvre, then a bit of GeForce2 GTS, and to crown all this with the GeForce2 Ultra dessert.
The heart of my long-awaited dessert...
Well, the restaurant itself was more than pompous both from outside and inside with various huge advertisements flashing all over the place - like 'The World's Fastest GPU' or 'MCPX'. Actually, it seemed to me that the place has become much bigger since my last visit there; so gigantic that I felt a bit uncomfortable inside. Anyhow, I found no strong-arm chucker-outs at the door, and this definitely helped my discomfort ease up a bit, which eventually perished when I was guided to my reserved table by a rather attractive and kind waitress (hmm...). I immediately ordered my three-course little dinner asking for the menu only to have a closer look at the ingredients of my long-awaited dessert. And I read the following:
- Second Generation Transform & Lighting Engines with the capacity of rendering 31 million triangles per second
- 250 MHz core frequency
- 64 MB 460 MHz 4ns DDR memory with 7.36GB/Sec Memory Bandwidth
- 4 independent rendering pipelines with 1 billion pixels/sec or 2 million texels/sec fill rate
- NVIDIA Shading Rasterizer (NSR) capable of per-pixel shading operations in a single pass
- High-Definition Video Processor (HDVP)
- AGP 4X/2X with AGP Texturing and Fast Writes
- Unified Driver Architecture
- 32-bit Z/Stencil Buffer
- Microsoft DirectX 7 and OpenGL Optimizations and Support
[oldal:The Dessert Anti-Aliased]I must admit that I did not devote enough time to the high-resolution textures and vibrant colors of my hors-d' oeuvre and GTS, and ate them greedily - eagerly waiting for the highlight of my night here as well as of NVIDIA's culinary masterpiece. And at last it arrived in all of its pomp:
The cream of NVIDIA cuisine
Well, looking at this appetizing spectacle, I did not want to break this harmony, and accordingly decided not to remove those aromatic bright green heatsinks to scrutinize what brand of 4 ns DDR SDRAM lay under them. Instead, I began to tenderly palate the Ultra, first to see what that controversial spice of FSAA can offer to the lovers of fine foods. Here I will make my best to illustrate what I managed to find out:
Q3:A at 640x480x32 with 2x2 (LOD bias) FSAA enabled
Q3:A at 1280x1024x32 with FSAA disabled
As it has become widely known by now, NVIDIA chips use supersampling to process Anti-Aliased scenes. Basically, this means that the chip renders the actual scene at a higher resolution, and scales it back down to the set resolution. And even though the procedure is hardware assisted, I had to observe a serious decrease in performance - ranging between 40 to 50 per cent depending on the resolution. Performance-wise it meant that I measured almost the same results under OpenGL with FSAA (2x LOD bias) enabled at 640x480x32 as at 1280x1024x32 without any sort of FSAA. And as it is shown above, you can hardly, if at all, perceive any difference between an Anti-Aliased scene at 640x480 and a non-Anti-Aliased scene at 1280x1024 - while in the latter case frame rates are much higher than in the former. In view of this, I chose to prefer FSAA off to FSAA on, and came to the conclusion that this spice - in its given mixture - was not tasty enough yet to worth spoiling the heavenly sensuality of this really delicious course.
[oldal:Some More Bitter Taste]After discarding FSAA, I moved on with enjoying my meal. But again I found something to bite on. This time I set to a routine work trying to recover my tasting ability, and installed Cyberlink's PowerDVD (ver. 2.55) bundled under various cloaks with most NVIDIA chip based cards (e.g. ASUS, Hercules and Leadtek). But much to my surprise, I was bound to observe that the DVD-player software could not handle the GeForce2 Ultra chip's hardware Motion Compensation feature with the latest official Detonator 3 driver (6.31).
MC in use
Sadly, this symptom reappeared in the case of the whole range of the GeForce series. First I thought that this must be a driver issue as MC worked fine with pre-Detonator 3 driver releases. But then I made an attempt with WinDVD, with which everything ran as smoothly as possible. As the lack of hardware MC has a noticeable effect on the quality of DVD playback (as you can observe occasional frame-drops), I do trust that both Cyberlink Corporation and NVIDIA will make their best to fix this annoying issue as soon as possible.
What you see is what you get
As I usually do not miss the opportunity to touch meals with cutlery made of various precious metals, I changed my golden fork and knife for those made of silver. You can hardly imagine what a difference this can make. And sadly it did make a big difference. My silver cutlery being the AMD Irongate chipset, I put to test what it can add to the taste of the most recent NVIDIA cream. And, no surprise here, I had to find that this combination was still something that spoiled the party. Most probably for safety reasons, NVIDIA's Ultra set its AGP transfer speed to AGP 1x by default. I know it well that this is an old story, but - in view of the fact that under D3D you lose at least 10-25% of potential performance in AGP 1x mode as opposed to AGP 2x - I find it more than exasperating that neither AMD nor NVIDIA have done anything to fix this issue ever since it cropped up a year ago. Still you are in need of some third-party tweaking utilities (like PowerStrip or NVMax) to set AGP 2x mode with your AMD 751 Northbridge to boost up performance. Under OpenGL there is not that much of a difference; but this is quite understandable, given that the GeForce series do not support AGP texturing under this API.
And there was another thing that did more harm than good to the overall picture I formed about NVIDIA's newest - or to be precise, about the diversity of ways its supported features were made use of. Despite new driver releases and patches, some software developers' implementation of texture compression (S3TC) offered by NVIDIA GPUs under OpenGL is still annoyingly unsatisfactory. The use of compressed textures still produced apparent visual artifacts in Quake3: Arena, as shown below:
Another brick in the wall
Of course, in the case of the Ultra equipped with 64 MB of RAM, you may as well turn off S3TC without having to sacrifice performance on the altar of image quality. But 32 MB boards do have some problems with handling large textures in higher resolutions and color depths.
Compare the above picture with this Unreal Tournament scene:
And some more bricks looking definitely better
[oldal:The Tidbits]After such gastronomic attacks upon my palate, I eventually decided to eat up the whole thing so that I could get an overall view of its virtues and vices. These tools I used to find out all I that was possible about the secret recipe:
Test system 1
- ABIT SE6 motherboard
- Intel Pentium III 733 MHz CPU
- 128 MB SyncMax PC133 SDRAM
- Maxtor 52049H4 20 GB UATA/100 HDD
- Diamond MX300 sound card
- Panasonic 5x DVD-ROM
- LG Studioworks 77i
- MSI MS-6167 motherboard
- AMD Athlon 800 MHz CPU
- 128 MB SyncMax PC133 SDRAM
- Maxtor 52049H4 20 GB UATA/100 HDD
- Diamond MX300 sound card
- Panasonic 5x DVD-ROM
- LG Studioworks 77i
- ASUS V7100 Pure (GeForce MX 32 MB) - 175 MHz core and 166 MHz SDR memory clock
- NVIDIA GeForce2 GTS 32 MB - 200 MHz core and 166 MHz DDR memory clock
- NVIDIA GeForce2 Ultra 64 MB - 250 MHz core and 230 MHz DDR memory clock
- 3DMark2000 ver. 1.1 by MadOnion.com
- Quake3: Arena ver. 1.17 by id Software
- Experience demo by WXP3D
- NVIDIA Detonator 3 (6.31)
[oldal:Flavors under D3D - PIII 733]So, let's have a look at our 3DMark 2000 results*:
I must admit that it did surprise me how consistent results were yielded in 16 bit. Note that the Ultra suffers the first serious drop in performance at the resolution of 1280x1024x32 only. This means that this is the point where the Ultra board starts to run out of its performance reserves. These charts as well as those that will follow are very good illustrations of the fact how prudently NVIDIA managed to segment its various GPUs.
I added the visually stunning Experience demo to the benchmark programs only because it takes maximum use of NVIDIA GPUs' T&L engine. Here, as you can observe below, the raw fill rate and memory transfer speed differences between my three courses manifest themselves in a very clear-cut manner:
Here, I guess, no comment is needed. So let's move on to our Athlon system.
*Note: the best results are indicated in the charts with the various shades of the red color.
[oldal:Flavors under D3D - Athlon 800]3DMark 2000 results on the AMD system:
It was interesting to see how the raw horsepower of the more powerful Athlon CPU had its beneficial effect on the benchmark results. However, the handicap of the motherboard chipset I was bound to use apparently showed off at higher resolutions and especially in 32 bit. It can be observed that moving on to higher resolutions and bit depths caused a gradual decrease in performance as opposed to the relatively consistent results of the PIII system.
Running Experience yielded only one interesting result only: namely that the MX card performed worse with the Athlon system:
[oldal:Flavors under OpenGL - PIII 733]For testing OpenGL performance I used the inevitable Quake3: Arena timedemo (demo 001).*
No surprise here. The Ultra came out on top at all resolutions and color depths. There is only one thing worth noticing: at lower resolutions the GTS was a decent competitor right on the tail of its big brother.
*Note: All settings in both the Display and Game Options menus were set to max.
[oldal:Flavors under OpenGL - Athlon 800]Athlon performance in Quake3: Arena:
With max settings, 90+ fps results can be regarded as more than impressive; here the Athlon system beats the PIII with quite a margin. Also, note that at both 640x480x16 and 800x600x16 the GTS performed on a par with the Ultra.
Partly I fell short of my expectations that the brand-new Pentium III 733 MHz system with its 133 MHz FSB coupled with AGP 4x would perform at least on a par with the Athlon 800 system based on an obsolete and unsupported AMD 750 chipset. As, despite the Irongate issues mentioned above, the raw horsepower of the definitely more powerful Athlon CPU apparently yielded better results in almost every testing environment. Due to the more enhanced features of the i815E chipset, it was only in D3D and in higher resolutions and color depths that the Pentium III could beat the handicapped Athlon system. While, under OpenGL, the AMD system was the absolute winner.
[oldal:Verdict]The dining room itself - as I have said - made me feel a bit uncomfortable because of its constantly growing baroque interior; but the waiters were kind (and occasionally pretty as well) and did not hover or stare. I think it was a serious blunder on the part of the proprietors that they had made the place a non-smoking restaurant. What a shame that I had to smoke my post-dinner cigar in the street! But, after all, this was a real exotic dinner - even though there were some problems with the trimmings. Though true it may be that I would not have had to taste this much bitter had I used some VIA KX/KT 133 based motherboard as second cutlery. But sadly, being pressed by time, I had no chance to obtain one. Of course, there is no doubt about that an i820E and AMD-760 system would have made the best couple of cutlery; however, I think that my dinner as described above did have its moral:
Selecting those chips that perform beyond the average and putting some steep expensive and rare DDR memory modules beside them can hardly be regarded as a real technological breakthrough. However, knowing NVIDIA's marketing policy, launching the Ultra was an inevitable and prudent step in terms of market segmentation. The only question mark I would put behind the prospective release of the GeForce2 Pro chip, as I can hardly find that little spot among the GeForce2 series where it could get a foothold.
The Ultra, just like the whole range of GeForce2 products, supports all 3D features needed in today's and tomorrow's (i.e. about half a year) applications and games. In this respect, only ATi's Radeon can surpass it at the moment. However, the lack of EMBM support is far from being a great disadvantage, as NVIDIA's dot-product bump mapping is definitely a better solution for creating the illusion of bumpy surfaces.
The GeForce2 Ultra is definitely the fastest card available on the market today. The only reason I rated it 9 out of 10 was that the GTS did manage to chase it closely, sometimes (in low resolutions) even to perform on a par with it.
500 bucks for a video card is something really horrifying. However, given its excellent performance and 64 MB of 460 MHz DDR memory on board, I say it is at least understandable - even though very few of us can afford it.
To sum it up, all I can say is that everyone has to decide it for themselves whether some extra fpss worth this much of extra money. The Ultra is a real expensive dessert for real gourmets. For your money, you get the fastest and most widely supported card on the market. While in 3D the MX performs well up to 800x600, the GTS up to 1024x768, the Ultra beats them all and has to face the first real challenge at 1280x1024x32 only. So if you feel like it, enjoy your meal!