Having decided to launch a new graphics chip every six months at any rate and reminding the world of this intention every time the chance occurs, NVIDIA got into a very unpleasant situation this autumn. The thing is that NVIDIA got too much ahead of its competitors, because it kept enhancing the graphics chip architecture incredibly fast. Really, almost all its previous rivals now turned unable to introduce new products performing close to those by NVIDIA as often. Matrox, for instance, can't finish the work on it G800 for over half a year already. That's why this graphics card will hardly look attractive against the background of the competing products. S3 has ceased developing graphics chips for a while. 3dfx, which delayed the launching of its VSA-100 chip for half a year, will be able to boast a product faster than GeForce2 GTS only if the company launches a four-chip Voodoo5 6000. However, even though they first promised to start selling this card in July, it isn't shipping yet.

So, the only chip developer, which can somehow withstand NVIDIA's invasion is ATi. RADEON is the only one today to show performance at least comparable with that of NVIDIA GeForce2 GTS accelerator. However, as soon as there appeared new Detonator drivers (6.xx version), RADEON, which has been a performance leader for a short while, had to step back and to return the laurels to GeForce2 GTS. Taking into account that there is no threat to the graphics giant's position in the today's market, launching a new chip is not a vitally urgent matter. On the contrary, it may harm the sales of the existing chips making it harder for the companies to sell their products on these chips.

On the other hand however, the today's low-cost graphics cards will allow you to feel quite at home even if you set one of the highest resolutions, such as 1280x1024 and up. That's why increasing the fillrate is not the No. 1 priority for the graphics accelerator developers. They would consider it much more interesting to implement some super cool new features, which could allow game developers to create more life-like 3D environments. However, it could be possible only if the API supports these new features. In particular, everybody expects new DirectX version to come out by the end of the year. So, from the marketing viewpoint it seems rather reasonable for NVIDIA and the others to wait until new DirectX is released, so that they could have enough arguments to back up the announcement of their accelerators' new features.

However, a word spoken is past recalling, and in order to keep its promise NVIDIA had to introduce something new. The company was quite aware that announcing NV20 now would be not the cleverest move that's why they decided to kill two birds with one stone. Having overclocked the already existing product they managed to leave their competitors even farther behind, to satisfy the customers waiting for something new when the 6 months expire, and to postpone its killer NV20 till a more favorable time.

GeForce2 Ultra Chip

This was just a short introduction. Now we would like to finally reveal our today's topic: NVIDIA's recent solution aka GeForce2 Ultra. As we have already said, NVIDIA resorted to overclocking methods. That's why its newcomer differs from GeForce2 GTS by two things:

• The chip frequency of GeForce2 Ultra got increased from 200MHz up to 250MHz;
• The memory frequency got increased from 333MHz (166MHz DDR) up to 460MHz (230MHz DDR).

In other words, NVIDIA simply overclocked its previous chip. In fact, we have to stress that GeForce2 GTS is simply an overclocked GeForce256 with a few improvements added, such as those dealing with Pixel Shaders, and with a doubled number of texturing units per pipeline. Other than that there are no new technologies. But, NVIDIA seems not to care about it at all and they go on using "brute force" methods to win the competition.

However, the situation is very likely to change with the launching of NV20. According to the already available info on this chip, it will be very much different from its predecessors mostly due to new technologies used. These technologies will allow reducing the memory bus workload and hence will eliminate the main bottleneck of the today's graphics accelerators.

But let's not veer too far away from our today's piece, NVIDIA GeForce2 Ultra, and try to find out why NVIDIA decided not to launch its Ultra chip together with GeForce2 GTS or not to make the working frequencies of GeForce2 GTS somewhat higher from the very beginning. As for NVIDIA, they claim that they are using an enhanced manufacturing technology for their Ultra chips. However, our experience in graphics cards overclocking proves that even the GeForce2 GTS graphics chips can be overclocked up to 250MHz, especially those launched lately. That's why we think that NVIDIA probably selects the most overclockable chips and marks them as Ultra chips. In this respect, we expect well overclockable chips to become a rare thing on conventional GeForce2 GTS graphics cards, because they will be used for more expensive Ultras. We have already witnessed the same situation with NVIDIA TNT2 and TNT2 Ultra chips manufactured simultaneously.

As far as the memory goes, 333MHz DDR SDRAM was the fastest memory available in mass quantities when NVIDIA launched its GeForce2 GTS. ATi RADEON, for instance, which came out a bit later was equipped with a faster 366MHz DDR. And now the market is ready to offer in mass new 4ns DDR SDRAM and NVIDIA simply can't miss this super opportunity. So, this memory is utilized in NVIDIA graphics cards aimed at extreme overclockers. However, 4ns memory can be also clocked to 500MHz without much difficulty, which NVIDIA was definitely aware of. Nevertheless, the company decided not to squeeze the maximum out of this memory and even reduced the memory chip frequency a bit, in order to make the product more stable. But on the other hand, this could be just a clever marketing move aimed at attracting the customers' attention by offering "a well-overclockable product".

Well, since NVIDIA is using the best, specially selected parts for its GeForce2 Ultra, the product appears quite expensive, which undoubtedly tells on the graphics card cost. When the first cards on NVIDIA Ultra chipset were announced they were supposed to cost about $500. However, hardly have two months passed and GeForce2 Ultra based cards started selling for $450. Is that a lot or not? To tell the truth, this cost seems quite understandable, especially against the background of a 350-dollar GeForce2 GTS with 64MB memory. Besides, GeForce2 Ultra is aimed at dedicated overclockers, who won't spare time and effort trying to find "the world's fastest graphics accelerator" and will definitely pay any crazy money for it. Moreover, the only product, which could compete with Ultra in terms of performance, is 3dfx Voodoo5 6000, which should cost $100 more if it ever appears in retail.

Now let's take a closer look at chip specs and compare them to the competing solutions available now:

GeForce2 Ultra architecture is just the same as that of GeForce2 GTS. There are four rendering pipelines, each featuring two texturing units, so that GeForce2 Ultra can process two pixels per clock in single-texture mode as well as in multitexturing mode. So, it turns out that in single-texture mode only 50% of all texturing units appear involved.

Due to the four pipelines together with pretty high chip working frequency NVIDIA GeForce2 Ultra can boast the today's highest fillrate in multitexturing: 1Gpixel per second.

However, it is mainly not the fillrate provided by the graphics processor that influences the performance of the entire graphics subsystem but the memory bus bandwidth, which is currently the gravest bottleneck of all graphics cards. From this point of view, 3dfx Voodoo5 6000 looks a much better balanced solution because the relation between its fillrate and the memory bus bandwidth is the most optimal of all. In fact, it may even appear faster than GeForce2 Ultra (if it comes out at all), but we won't state anything for sure now. Even though the memory bus bandwidth by GeForce2 Ultra is the highest of all among graphics accelerators based on standard architecture, it remains the main bottleneck, nevertheless. If only NVIDIA introduced some technologies to reduce the memory bus workload, then…

Closer Look

Well, this time we managed to get a graphics card from Creative based on NVIDIA GeForce2 Ultra chip. It was Creative 3D Blaster Annihilator 2 Ultra card.

Since this time Creative again was the one to develop the card reference design for NVIDIA (they also worked on the GeForce2 GTS reference card), the card we had at our disposal was an absolute copy of the reference card. It was equipped with 64MB DDR SDRAM, however, it had no TV-Out and no DVI-Out though the PCB layout implied the use of both. Although these options were included into the reference design, Creative made up its mind not to provide its graphics card with them. Taking into account that the users will purchase the 400-dollar graphics cards for playing games at higher resolutions and that TV-Out and DVI-Out do not support anything over 1024x768, Creative decided to save time and trouble and to design the card without them. Of course, this approach allowed Creative to make its product considerably cheaper than the graphics cards on NVIDIA GeForce2 Ultra from other manufacturers.

The memory used on Creative 3D Blaster Annihilator 2 Ultra is a bit different from what you can see on NVIDIA reference card. Unlike ESMT memory chips on the reference card, our memory chips were made by EliteMT and featured 4ns access time. To tell the truth, the nominal working frequency for these memory chips makes 500MHz and 4.5ns memory could be just enough for GeForce2 Ultra with the 460MHz nominal frequency. However, NVIDIA, which supplies its graphics chips together with the memory chips to the graphics card makers decided to eliminate every single risk that something might not work properly and chose the memory with a bit higher resource. Besides, high EMI, which appears at high working frequencies like in Ultra's case and worsens the card stability, also may have pushed NVIDIA towards this decision.

Besides, in order to ensure stable and reliable performance, the memory chips were provided with heatsinks like those Hercules used on GeForce2 GTS based cards. However, we wouldn't say that the memory on Creative 3D Blaster Annihilator 2 Ultra heats too much. So, the primary role of the heatsinks fastened to the memory chips seems to be the protection against EMI and not the heat dissipation. Proving this point, we may say that with the heatsinks removed, the graphics card was much less stable, while the memory chips got only 2-3 degrees warmer.

Moreover, there is one more heatsink with a fan set onto the graphics chip. Even though this cooler is quite small, it provides a rather strong air stream, which ensures the required cooling of the chip working at 250MHz.

By the way, when we removed the cooler from the card, we found out that the GeForce2 Ultra chip on our Creative 3D Blaster Annihilator 2 Ultra had a metal lid, while GeForce2 GTS chips were covered with plastic lids. This modification definitely allowed to improve the heat dissipation and the chip cooling and to protect the latter against increased EMI.

Creative 3D Blaster Annihilator 2 Ultra is accompanied by the drivers based on NVIDIA reference driver version 6.25, which include the company's BlasterControl utility. It allows accessing the driver settings via a small icon located in the system tray.

Although this utility looks somewhat extravagant, it contains all the settings of the reference drivers.

Well, enough theory. Let's pass over to the most interesting part of our review: the tests.

Performance

For our tests we assembled the following testbed:

• Intel Pentium III 866 CPU;
• ASUS CUSL2 mainboard;
• IBM DTLA 305020 HDD;
• 256MB PC133 SDRAM by Hyundai (2 x 128MB);
• Windows 98.

For a more illustrative comparison we suggest also taking a look at the performance of the following accelerators:

• NVIDIA GeForce 2 GTS with 64MB DDR SDRAM: ABIT Siluro GF256 GTS (NVIDIA Detonator v. 6.35)
• NVIDIA GeForce 2 Pro: ASUS V7700 Deluxe 64MB overclocked up to 200/400MHz (NVIDIA Detonator v. 6.35)
• ATI RADEON 64MB DDR (driver v. 1.0305)
• 3dfx Voodoo5 5500 (driver v. 1.05)

We will use the following two games to consider the performance of Creative 3D Blaster Annihilator GeForce2 Ultra in 3D-graphics:

• Quake3 Arena 1.16n. We ran demo001 and quaver demos to test the cards in OpenGL.
• Unreal Tournament 4.32. We used utbench demo to test the cards in DirectX.

Before we start, we would like to draw your attention to the fact that we didn't test the graphics cards in applications supporting HW T&L, because activating HW T&L implemented in some of the today's games doesn't provide any tangible performance gain. That's why the usefulness of the T&L engine now can be argued about, actually. However, we do not deny that this feature as well as Pixel Shaders will be soon demanded.

So, here we go.

As we have expected, GeForce2 Ultra proved the fastest in 16bit color. No wonder, since it is an overclocked GeForce2 GTS, which has always been the leader in this mode.

In 32bit color, GeForce2 Pro and GeForce2 GTS fell slightly behind their elder brother. While ATi RADEON and 3dfx Voodoo5 5500 proved considerably slower than GeForce2 Ultra. Voodoo5 5500 turned out twice as slow and RADEON - 1.5 times slower, which is quite noticeable.

At first sight, everything seems to be pretty clear in this benchmark. GeForce2 Ultra proved up to our expectations and left everyone far behind. Nevertheless, we suggest taking a closer look at some peculiar details of the new NVIDIA creature before we continue praising it. Is everything as good as it seems at first?

The first thing that draws our attention is the fact that the graphics card is surprisingly ill-balanced. Of course, we never doubted the power of the graphics chip featuring four pipelines with two texturing units each. However, when we tried to find out its real potential we learned that sometimes it hardly uses 50% of its resource!

What made us think so? Well, here is a bit of easy math1ematics to back up our point. Let's take the resolution 1024x768 in 32bit color as a starting point for our investigation. This is the resolution when the memory bus bandwidth starts proving insufficient for the powerful GeForce2 Ultra graphics chip. (Note that all the calculations are empirical and the real graphics card performance may appear higher or lower than the obtained values).

The "effective" memory bus bandwidth (disregarding overdraw, z-buffer, etc.) at the resolution of 1024x768x32 equals: 1024*768*4(32bit)*115.6fps (taken from the diagram above)=346MB/sec. The memory bandwidth is the only thing that tells negatively on the graphics card performance. The chip fillrate cannot be a limiting factor because it makes only around 90Mpixels in this resolution (1024*768*115.6), which is quite far away from the announced fillrate of 1000Mpixels. Moreover, even at the higher resolution, take for instance 1600x1400x32 with the same 115.6fps, the chip fillrate will still remain out of the question: 1600*1200*115.6=220Mpixels. In our tests we got only 55.4fps. In other words, everything proves that only the graphics memory is too slow for such a fast chip.

And now let's try to find out which memory should we take for the system built with Pentium III 866MHz CPU, so that it didn't limit the system performance (provided we have 128fps at all resolutions).

The required "effective" bandwidth will make 1600*1200*4*128=938MB/sec. From a very simple proportion we may get the memory frequency: 460MHz*(938/346)=1247MHz. In other words, we need a twice as fast DDR SDRAM, which will hardly every be developed.

So, judging by everything mentioned above we may draw two very important conclusions:

1. Increasing the working frequency of the graphics chips with classical architecture won't make any sense if the memory remains the same as we have now. What's the use of a McLaren car on a village road?
2. The CPU may appear not powerful enough for the graphics accelerators like GeForce2 Ultra, for instance. This statement can be illustrated by the following example.

The "effective" graphics memory bandwidth allows achieving 189fps at the resolution set to 800x600x32bit: 346/(800*600*4)=189fps. However, in Pentium III 866MHz built system GeForce2 Ultra shows only 128fps. If we carry out some simple calculations, we will get that in order for the graphics card to achieve its maximum at this resolution, the Pentium III CPU should work at 866MHz*(189/128)=1279MHz.

As is known, there will be no Pentium III CPU supporting the frequency over 1.13MHz that's why we don't have the chance to see the maximum GeForce2 Ultra is capable of with these processors. And as soon as the first faster CPUs turn up (such as Athlon, for instance), there will be a new NVIDIA chip coming: NV20. Therefore, we would call NVIDIA GeForce2 Ultra based cards not only the absolute leader of the graphics marathon but also the card of unrealized potential…

Well, let's return to our tests and continue our investigation aimed at finding out how much fast GeForce2 Ultra is.

In quaver demo program in 16bit color mode GeForce2 family left its competitors far behind. At lower resolutions (even 1024x768 is now regarded as a low resolution for these graphics cards) they perform almost equally, the difference makes only 1-2fps, while at higher resolutions GeForce2 Ultra makes a significant dash forward and leaves behind not only 3dfx Voodoo5 5500 and ATi RADEON, but also GeForce2 GTS and GeForce2 Pro. Besides, it is also worth mentioning that at higher resolutions ATi RADEON proved the slowest of all, even slower than 3dfx Voodoo5 5500.

In 32bit color the situation didn't change at all. GeForce2 Ultra held the leading position at all resolutions. The performance difference was especially noticeable at 1600x1200, where even GeForce2 Pro lagged 19% behind the Ultra card. Not to mention all the other products. Now you see what the price of the highest graphics memory bandwidth is!

The results in Unreal Tournament turned out pretty interesting. All the cards except Voodoo5 5500 turned out almost equal in both 16bit and 32bit color modes.

Full-Scene Anti-Aliasing

Since we have already seen that NVIDIA GeForce2 Ultra provides quite high fps rates in the today's games, enabling FSAA finally appears possible without losing too much fastness. By the way, the 64MB memory our today's hero is equipped with allows using FSAA 2x2 in all resolutions except the highest: 1600x1200x32:

As you can notice from the diagrams above, 2x2 FSAA can be easily enabled for 640x480 and 800x600, and 1.5x1.5 FSAA also suits for higher resolutions. But nevertheless, if the user is buying this relatively expensive graphics product, GeForce2 Ultra based card, he or she is most likely to have a large monitor and hence will prefer setting 1600x1200 resolution to enabling FSAA.

Overclocking

As for overclocking, Creative 3D Blaster Annihilator 2 Ultra featured everything necessary for that: the graphics card we had at our disposal was equipped with 4ns graphics memory potentially supporting 500MHz, the memory chips were provided with heatsinks and the specially selected GeForce2 Ultra chip was covered with a metal lid. So, we managed to overclock our card up to 315MHz core and 500MHz memory frequency. In fact, we expected this significant gain (the chip frequency got 25% higher and that of the memory - 10% higher) to tell positively on the performance. Here is what we got:

As you see, there is hardly any performance gain at all at 1024x768x16bit. It means that even our Pentium III 866MHz couldn't provide sufficient workload for the powerful GeForce2 Ultra based card. As far as higher resolutions are concerned, the maximum performance gain achieved in 16bit color mode due to overclocking was obtained at 1600x1200 and made about 10%. Note that the effect made by the 25% chip frequency growth is just the same as that made by the 10% memory frequency growth.

In 32bit color overclocking has more influence on the cards performance. Note that increasing the graphics memory frequency by 10% provides the performance gain twice as high as in case of chip frequency increase by 25%. This is another argument for the fact that graphics memory bandwidth is the major factor influencing the performance of NVIDIA GeForce2 Ultra in the first place. As for the general performance gain we get, it makes 7fps (13%) at 1600x1200x32.

And now a few words about Creative 3D Blaster Annihilator 2 Ultra working at the non-nominal AGP frequencies. Really, those of you who are ready to purchase the today's fastest graphics accelerator may be pretty interested in using it with the i440BX based mainboards, which is considered the today's fastest chipset (133MHz or 150MHz FSB). That's why we tried to make our GeForce2 Ultra work on a system built with ASUS P3B-F mainboard based on this chipset and featuring the FSB overclocked up to 133 or 150MHz. In the first case, the card ran perfectly stable, however, with the 150MHz FSB the system kept crashing quite unexpectedly.

Conclusions

Well, NVIDIA GeForce2 Ultra can't boast any architectural enhancements compared to its predecessor, GeForce2 GTS, and is none other but an "officially" overclocked GTS chip. However, GeForce2 Ultra proves nevertheless much faster than any of its today's competitors. 3dfx Voodoo5 6000 hardly has any chance to surpass its new rival. Besides, it seems to us that the more chips there are on the graphics card the less efficiently it works. Moreover, 3dfx keeps postponing the launching of Voodoo5 6000. Even if NV20 doesn't come out before next spring, GeForce2 Ultra will be able to retain the leadership for a considerable while.

To tell the truth, we don't think that spending the whole bunch of money on a GeForce2 Ultra based graphics card would be justified now. In fact, there will be some other more up-to-date solutions available by the time the fillrate provided by GeForce2 Ultra is used to the full extent.

However, no matter what, there will be people willing to by this cool piece even for the crazy money it costs. Hardcore gamers, hardware collectors and exotic lovers may make the demand for Ultra products.