by FastSite
04/30/2001 | 12:00 AM
NVIDIA GeForce3 chip was announced in February, to be exact - on February 27, 2001. Two months have already passed since then, and it is only now that the first graphics cards built on GeForce3 start cropping up in retail. The delay was caused by some bugs found in the GeForce3 chips. No wonder, since the core of GeForce3 comprises more than 50 million transistors and it's as complicated as modern CPUs by AMD and Intel. It took some time to eliminate all the bugs, so the final revision is A05, i.e., only the fifth chip version was considered suitable for mass production. This situation is not typical of NVIDIA. For example, its ultimate GeForce256 and GeForce2 chips featured revision A03. All the cards we tested in our previous article called NVIDIA GeForce3 Guide were equipped with A03 chip revisions, which were not the final chips versions, as we found out now. As for the "ready-to-serve" A05 chips, they started shipping to the graphics cards manufacturers in late March. And the graphics cards based on them will start selling by almost all the well known manufacturers in May already.<%BANNER[article]%>
We managed to test a mass piece based on the final revision of GeForce3 chip - VisionTek GeForce3 by VisionTek Inc., which was one of the first to start mass production of graphics cards built on the A05 chip revision.
VisionTek GeForce3 graphics card based on NVIDIA GeForce3 chip was in the spotlight of our investigation. Here it is:
The card is shipped in retail package. Unfortunately, our package contained only the card and a user's manual.
As you can see on the photos, the card is made according to NVIDIA's reference design. VisionTek GeForce3 is bundled with a S-Video Out, and TV signals encoding is carried out by Bt869KRF chip.
The chip of VisionTek GeForce3 works at 200MHz. It is marked not as "NV20", but as "GeForce3" and has A05 revision, as we have already told you, thus it means that we deal with the final chip version, which is installed on the mass graphics cards. This chip is manufactured on the 13th week of 2001, which indicates that NVIDIA began to ship the mass chips to the manufacturers in late March or early April.
This graphics card is equipped with 64MB 128-bit memory by EliteMT clocked at 230MHz (460MHz DDR) with 3.8ns access time:
Graphics memory like that should theoretically be able to perform well at the frequencies up to 530MHz, but according to NVIDIA's recommendations its working frequency is set to 460MHz. This allows increasing the system's stability and reliability and preventing EMI. At the same time, the memory is supposed to overclock quite well.
The chip is equipped with a cooler, which is usually installed on reference graphics cards based on GeForce2 Ultra and GeForce3 by NVIDIA. The memory chips are covered with ribbed heatsinks, which get sufficient ventilation from the cooler thanks to their shape and the shape of the core heatsink.
Our testing system was configured as follows:
We tested with the following software:
Drivers:
We tested VisionTek GeForce3 card in non-nominal modes without any supplementary cooling. All the tests were run in an open testbed at approximately 18-22 C room temperature.
We've found out that the highest working frequencies for VisionTek GeForce3 when the card could keep working are 265MHz for the core and 385MHz (570MHz DDR) for the memory. But in this mode the card's performance was not stable enough: in 10-15 minutes after starting Quake3 either some annoying artifacts popped up, or the system simply crashed.
By extreme over-overclocking the hanging was "due", that is, the image froze as it should happen with all smart graphics cards when they are overclocked :-) We observed no artifacts, which were mentioned in our previous article dedicated to GeForce3 (we mean the evenly distributed rectangles).
In order to complete the 3Dmark2001 tests, we brought the frequencies of the core down to 240MHz and the graphics memory - down to 270MHz (540MHz DDR). Otherwise the card failed to withstand the whole set of tests lasting about 40 minutes.
We can admit that VisionTek GeForce3 is good to clock up to 240/540MHz, when the core frequency is 20% higher than the nominal and the memory frequency is 17% higher. These figures are not bad at all, especially bearing in mind that the fastest graphics cards of the previous generation - those based on GeForce2 Ultra, couldn't pride themselves of overclocking potential that high.
In Quake3 Arena we set the textures used to the maximum and enabled the highest level of detail and trilinear filtering. Other settings were left as default. We tested in demo127.dem included in 127g point release patch for Quake3 Arena.
Tests in 3Dmark2001 were run with default settings, in 16-bit color mode we enabled 16-bit textures and 16-bit Z-buffer depth, while in 32-bit mode we had 32-bit textures and 24-bit Z-buffer.
We decided not to use Unreal Tournament as long as all GeForce3 cards (even overclocked and run in utterly high resolution modes like 2048x1536) showed extraordinarily similar results. We didn't dare to draw any conclusions basing on them. It is undoubtedly the engine of Unreal Tournament, which is to blame here, because its performance strongly depends on the CPU and the system memory and is hardly influenced by the graphics card itself.
Alongside with VisionTek GeForce3, for a better comparison we took a pre-production sample of PixelView XX-Player graphics card by Prolink based on NVIDIA NV20 with A03 stepping.
A third racer was Creative 3D Blaster Annihilator 2 Ultra based on NVIDIA GeForce2 Ultra.
At nominal frequencies the performance of VisionTek GeForce3 was very close to what we have already shown in our NVIDIA GeForce3 Guide. This fact as well as the splendid overclockability of VisionTek GeForce3 made us pay special attention to analyzing this graphics card from the point of view of overclocking.
For this purpose, we tested VisionTek GeForce3 in four modes: at nominal frequencies, with an overclocked core, with the overclocked graphics memory and after clocking up both the core and memory.
Naturally, in 16-bit color mode the graphics memory is not loaded too heavily, therefore GeForce3 core yields a notable performance gain when it works at over-nominal frequencies. In 32-bit mode the graphics memory bandwidth is supposed to become a sticky point, but by overclocking the core and the memory in this mode we got almost the same upturn in performance. The developers have obviously done a lot to balance GeForce3 chip, and Quake3 lets us enjoy in full the result of their work. To illustrate the influence of the overclocked core and memory on the performance in Quake3, we have drawn some graphs for you. The y-axis of the graphs below shows the performance growth in case of overclocked core, memory and core-and-memory frequencies.
It's pretty logical that at 800x600 and partially 1024x768 resolutions the performance is limited not by the graphics card, but by the performance of the CPU and the system memory. That's why we don't see any striking performance increase at these resolutions. In 16-bit mode the overclocked memory proved ineffective. The performance was determined by the core frequency, but starting from 1600x1200 the graphics memory bandwidth makes this effect almost nil. United efforts of the overclocked core and memory bring about an impressive 20% general increase in performance. The core in this case gets clocked up by 20% (from 200MHz to 240MHz) and the memory - by 17% (from 460MHz to 540MHz).
In 32-bit color mode individual core and memory overclocking gave us more or less equal results, but at higher resolutions the effect was not so great. The thing is that although the memory is overclocked, the core, where HSR-related calculations take place and the caches are located, sticks to the same initial clock frequency. In a roundabout way, when the core is clocked up, at high resolutions the performance growth is restricted by insufficient memory bus bandwidth.
By parallel core and memory overclocking the total performance gain is nearly 30%. Marvelous, isn't it? Due to high-quality electronic bundles and simple but thorough cooling, VisionTek GeForce3 feels well at over-nominal frequencies without any supplementary ventilation.
It may seem shocking, but when only the core or only the memory of GeForce3 get overclocked, there appeared absolutely no positive effect. The performance gain lies within the interval stading for possible measurements inaccuracy. However, by simultaneous core-and-memory overclocking there is a real performance gain (from 10 to 20 percent depending on the resolution). Why so? As we know, GeForce3 architecture is really well-balanced. Then, 3Dmark2001 itself is optimized for GeForce3. If geometric calculations and texturing load tangibly both the core and memory, too slow data transfer becomes a bottleneck as core frequency increases. And vice versa - when a clocked up memory overcomes the problem of slow data transfer, geometric calculations are done no faster. The chip starts texturing the triangles no sooner than it receives vertex data from the T&L unit. Hereat the performance of the geometrical unit becomes a foible as the memory is overclocked. Only when we increase both core and memory frequencies, an appreciable performance gain is achieved.
In High Details mode it turned out absolutely unreasonable to overclock GeForce3 at low resolutions. Apparently, even Athlon 1.2GHz can't cope with the physics calculations included in this test and load GeForce3 with the data. Some gain takes place only in 1600x1200 resolution.
Everything we commented on in Car Chase test is true for this case as well. The only principle difference is that in this test the CPU calculates no physics, leaving the graphics card responsible for the performance in all resolutions. Simultaneous overclocking of the core and the memory results in a 15-20% performance increase depending on the resolution.
The performance of VisionTek GeForce3 at nominal frequencies, here as well as in most of the tests described above, comes abreast or a bit ahead of the reference cards performance, which we discussed in our GeForce3 Guide. It means that the new revision of the chip differs very slightly from its predecessors from the performance point of view.
At low resolutions the performance of our hero in this test is seriously restrained by the system speed: in this case the overclocking of GeForce3 brings no visible effect. In this test the number of polygons and textures is two or three times smaller than in the previous tests. Besides, Vertex Shaders deal only with shadows and characters skinning. That's why the results are very much like those in Quake3 Arena - single memory and core overclocking is fruitful, but in 16-bit mode it's much better to clock up the core, while in 32-bit mode it makes no difference.
Here we encounter active use of Vertex and Pixel Shaders, so the performance is much determined by the fillrate and the fastness of geometrical calculations.
The results confirm our assumptions: in both 16 and 32-bit color modes we get higher performance gain by overclocking the core rather than by overclocking the graphics memory.
In single-texturing mode the results strictly correspond to the theory - in 16-bit mode the fillrate is confined with the core frequency and grows more notably when we clock up the core. And in 32-bit mode the performance depends only on the memory and goes up together with the growing memory frequency.
In multi-texturing mode we were perplexed to see core overclocking being more effective than memory overclocking in 32-bit color mode. Especially since in this fillrate test the memory experiences the hardest strain of all.
We'll try to give some explanation of these results. You see, in this test there is only one semi-transparent texture, which is laid over the polygons. Moving-stripes effect is achieved by simply moving the texture when laying. Of course, when a texel from one of the texturing units is requested, GeForce3 caches part of this texture and perhaps some texels for other units. As the core frequency grows, the cache data is processed quicker.
These synthetic tests check how fast GeForce3 core transforms the vertexes and copes with Vertex and Pixel Shaders. No surprising trends were discovered: as it had been anticipated, the performance grew by core overclocking. GeForce2 doesn't support Vertex and Pixel Shaders, so for this chip we had to make CPU emulate Vertex Shaders and to abandon Pixel Shader test.
As we can deduce from Aquanox, overclocking GeForce3 is also reasonable for the upcoming games. Unfortunately, the developers give no details about the engine, and we lack the data for the analysis of the results obtained for memory overclocking and for core overclocking, that's why we provided the results obtained only for simultaneous core-and-memory overclocking.
VisionTek GeForce3 graphics card revealed remarkable stability, performance and overclockability. It should be noted that it's one of the cheapest GeForce3 cards - from the very first day it should start selling for $370.
Having tested VisionTek GeForce3 with the ultimate NVIDIA GeForce3 core, we made sure once again that GeForce3 is perfectly balanced and has a great potential.
As our tests showed, VisionTek GeForce3 doesn't always bring due performance gain if only the core or only the memory get overclock. In this case the performance in 3Dmark2001 doesn't rise that much, which means that you won't get the expected gain in the upcoming games as well. On the other hand, in Quake3 we saw that in the already existing games GeForce3 core is still worth overclocking. Anyway, the situation must become clearer as the developers release new test applications and games, which will make the most of GeForce3's capabilities.
Highs:
Lows: