GIGABYTE 3D1-68GT: Double fun

Preface

We first heard of Gigabyte's 3D1 video cards in December last year. 3D1-68GT continues the glorious tradition of dual-chip solutions. But before proceed directly with the 3D1-68GT, let's bring back into history for a while.

The idea of harnessing two GPUs in one team is not new in itself. More or less successful attempts have been made for quite a long time. The passed away ambitious 3dfx used to produce video cards with 2 and even 4 VPUs onboard. ATI's Rage Fury MAXX also used to offer two VPUs onboard. So what do we see now? As you know, NVIDIA having acquired 3dfx with all of its developments somehow reworked the idea of merging the performance of two VPUs and suggested fitting two separate video cards running in combination into a single system. Such approach gained recognition from its major competitor. Of course, I mean ATI with its CrossFire technology.

As some classic once noted, history progresses in spiral. While creating their first 3D1, engineers at Gigabyte inherited the superficial idea and merged two 6600GT video chips on a single PCB. In the end, they produced a video card of performance superior at some applications over the leaders of that time.

Despite the saying that you must spoil before you spin, the very first try proved successful. Even some limitations specific of the 6600GT chip did not spoil the impression. Albeit the 128-bit memory bus did not allow showing the full at the most demanding modes with the FSAA and AF enabled, that is not so big. What matters is that through the release of 3D1 not only did Gigabyte paved a way to new solutions based on standard VPUs - it also broke new grounds for fantasies like merging two dual-chip video cards in a system thus making 4 GPUs.

The choice of 6600GT as the basis for the first 3D1 was justified. Creation of PCB to host two VPUs on, the interface between them, own video memory for each GPU, complicated power supply system - all these took great efforts.

GIGABYTE 3D1-68GT

We only have to conjecture what sort of difficulties engineers at Gigabyte faced while creating 3D1-68GT. Nevertheless, they didn't give themselves an easy time and merged two full-featured GeForce 6800GT on a single PCB with as much as 512 MB of video memory - 256 MB to each video chip and a 256-bit bus per each chip. They enhanced the power supply system, fastened a powerful cooling system and raised the frequencies relative to the standard.

Let's see what has proved of these all.

For viewing convenience, one photo shows both the front and rear views of the box. The box is designed in tender blue tints. On the front side of the package (photo to the right), instead of the traditional fairy-tale heroes or extraterrestrial monsters, there is the inscription "3D1" set off in wreath and ribbon. Sort of a modest hint for superiority. Of course, the wreath is not given for the second place. : The inscription "2005" on the ribbon hints that next year we may see another solution of the 3D1 series. Below are the main features of 3D1-68GT, as well as the results of benchmarks (of course, 3DMark), and photos of the video card itself.

On the reverse side, central part is taken by the photo of 3D1-68GT with the detailed description of all the features and capabilities of the video card. Here are also detailed specifications (to be disclosed in depth in what follows).

Inside the box, we found the following:

• User's Guide
• CD including the video card drivers, DirectX, PowerDVD 5.0, and the proprietary utility for overclocking video cards - V-Tuner2.
• CDs with games: XpandRally and Joint Operations
• Special strap to connect two additional D-SUB connectors
• Two DVI/D-SUB adapters
• Adapter for additional power supply
• Adapter for TV-out S-Video/Composite

They might have put more game CDs, and I wish the games were more modern. The card pretends to offer the top-end performance level and will cost a lot. On the other hand, for Russia that is not a critical issue due to known reasons.

…and of course the video card itself

The photo in this perspective can't express the volume, unfortunately. And no photo can express the weight of the thing. Believe me - 3D1-68GT is spacious enough and weighs quite substantially. First of all, because of the massive cooling system which hides the PCB almost completely from both sides.

Cooling system.

A few more photos.

Front view, in a different perspective. See the height of the radiator. The height is so large that it can easily fit a pretty big cooler. As you understand, the neighboring PCI slot will be overlapped completely. It is not recommended to install something into the second PCI slot. You can inadvertently block the air inflow to the cooler. The radiator cover with a proprietary label is needed not for nicety alone. It serves to direct air flow along the radiator fins that run in parallel to the longer side of the video card over the whole length.

Front view with the lid removed. As can be seen from the photo, the radiator installed on the front side of the video card is made up of two parts. Why do you think it is so? The answer is evident - to provide a thorough pressing of the radiator to each of the GPUs. With a solid radiator, that would have been very difficult to do. Since there is only a thin layer of thermal paste between the GPU and the radiator, a very precise alignment of the VPUs in height and plane would be needed, or the thermal paste would have to be replaced with the so-called "thermo chewing-gum", which would substantially deteriorate the cooling mode of video chips already hot enough.

The rear view. As you can see, a cooler is there on the reverse side either. The increased video memory capacity is to blame for that. The radiator's dimensions (for video memory alone!) are indeed enviable for any video card of the middle price range. Such solution looks a bit redundant but is logical enough in terms of the general concept of building a cooling system.

The radiator that closes power elements of the supply system in this regard almost interflows with the main radiator. The concept has touched upon it either.

This is how the radiators look from the reverse side when removed. All the surfaces are polished almost to the mirror gleam and leave the most favorable condition. But that doesn't play a big part for the memory chips because thermo-gums are used as the thermo interface. As can be seen from the photo, radiators are fastened not to the PCB but to one another thus forming sort of a "bread-and-butter" inside which the video card is fitted into.

Let's see what is hidden under this massive splendor.

See how tightly the components are positioned. The seemingly empty spots are in fact cobwebs of tracks leading to the memory chips. As you can see, the second VPU is turned 45 degrees relative to the first one. You might expect that the overall pattern of positioning of the video memory chips relative to the GPU will be preserved and turned in a similar way - nevertheless, that is not the case. A simply turn would require increasing the PCB size. Engineers at Gigabyte decided not to look for easy ways and demonstrated miracles of resourcefulness.

This is how one of the two GPUs installed onboard looks. It was produced on the 13th week of year 2005. Don't be confused by the number of production week. Running ahead, I'd rather say that during the tests both processors showed themselves more than with dignity.

The memory made by Samsung offers 1.6 ns access time, which makes maximum 1200 MHz DDR operating speed. At the same time, the nominal frequency of the video memory for 3D1-68GT is 1000 MHz. Such a margin for video memory frequency lets one hope for a substantial performance boost of the video subsystem.

On the reverse side of the video card, the picture is fully identical - a lot of fine elements, dense assembly, another 8 memory chips.

At that, we'd rather stop scrutinizing the video card and move on to the benchmarks.

Specifications

For reference, we are bringing in the specifications for 3D1-68GT video card:

• Two GeForce 6800GT VPUs, operating frequency 375 MHz
• Total capacity of installed video memory - 512 MB, 16 chips in the BGA package, 8 per each side of the video card
• Two memory buses, each of 256 bit (each GPU uses its own 256 MB of memory)
• Memory type - DDR3, with 1.6 ns access time, operating frequency 1000 MHz
• Interface – PCI-E 16x

As you can see, the GPU frequencies are somehow overstated relative to the standard 350 MHz for 6800GT. It's hard to say what was the reason for such frequency overstatement. Perhaps the designed cooling system did such a good job that it was decided to raise the GPU operating frequency. Or, on the contrary, raise of frequency resulted in the need to develop such a massive cooling system. I presume the most probable reason was the manufacturer's wish to produce a result that would be much higher than the 20 000 marks threshold in 3DMark03. You might ask why 3DMark03 - it's now year 2005 which is coming to an end? We'll get the answer once we move to the tests. On the other hand, it is widely known that many video cards built on 6800GT easily overclock to frequencies equivalent to 6800Ultra. However, the "nominal overclocking" is a bit more modest. Among the natural limitations here is not much to do with the overclocking potentials of the GPU, but the increased total power consumption and heat emission of the video chip tandem.

The memory frequency fully meets the nominal for 6800GT even though the installed memory allows running at a higher frequency. On the whole, it us reasonable. Further raise of memory frequency does not produce such a radical effect upon the performance, but the above limitations for power consumption and heat emission play their part in here either.

However, despite all the reasoning behind the design and selected operation mode, it was not possible to avoid a fly in the ointment. As it turned out, the video card always runs at the same frequencies - both in the 3D and in the 2D modes. As you understand, such a behavior can't have a positive effect on the thermal mode of the video card and the system on the whole.

The second fly in the ointment is in that temperature monitoring is missing in 3D1-68GT. Neither RivaTuner, nor V-Tuner2 detected any temperature sensors, so showed nothing. Nor control for the coolers' rotational speeds was detected, which looks logical - the video card always heats up to the maximum.

It's hard to say what was the cause of mentioned flaws. Perhaps it is related to that engineers at Gigabyte had to write a special BIOS for 3D1-68GT, which resulted in some problems. Or, perhaps these are the tradeoffs for dual-chip nonstandard design. Or perhaps the lack of time. Anyway, that does not affect the stability of the video card. In any case, one will have to take care of a good ventilation of the housing and a powerful reliable PSU.

Overclocking

Prior to overclocking the video card manually, I decided to verify the automatic operation mode and "asked" the system to detect the optimum frequencies. The result surpassed my expectations. The drivers agreed that 375 MHz for the GPU is good albeit not enough and set 394 MHz, so decided not to touch the memory thus letting it merely 1030 MHz. Not bad. Recalling the overclocking capabilities of 6800 GT to the level of 6800 Ultra and the memory frequency margin, I certainly did not stop at that and decided to overclock 3D1-68GT a bit. In the end, the card showed a stable operation at 425/1200 MHz. As you can see, the frequencies meet the nominal frequencies for 6800 Ultra and even surpass them a bit (for the memory).

Benchmarking

Some preliminary notes (for those who don't like reading manuals). If we install 3D1-68GT into a motherboard that does not support SLI, the video card will run as a single 6800GT, but admit what's the point of having two GPUs then? So, to provide operation of 3D1-68GT video card in the maximum performance mode, a motherboard with support for SLI is a must! However, there are some fine points in here. Although support for SLI is a must, the switch changing the modes Normal/SLI has to be set to the Normal position. On the video card, the SW1 switch has to be set to the "ON" position (with the white dots turned to the bottom), as is seen on the Figure.

The switch itself is positioned on the rear side of the video card.

Attention! To change the SW1 switch to a different position, don't remove the radiator from the video card! The radiator won't close it.

Test-bench configuration:

For tests, we used NVIDIA's latest ForceWare 78.01 WHQL drivers.

3D1-68GT was tested in two modes. First - nominal frequencies 375/1000 MHz. Second - overclocked version, frequencies 425/1200 MHz (GPU/memory, respectively). The results were compared versus GeForce 7800GT, 7800GTX, and, since 3D1-68GT is essentially a couple of 6800GT in the SLI mode, we were curious to compare it against the top SLI-solution to date - 7800GTX-SLI.

Here are the results:

As expected, at the nominal frequencies 3D1-68GT takes a sure lead over 7800GT/GTX and approached the 10 000 marks "threshold". Overclocking helped to get over the 10 000 marks bar! Overclocked 3D1-68GT even surpasses 6800 Ultra-SLI perhaps due to a bit higher frequency of the video memory. Sure, 7800GTX-SLI remains inaccessible.

The alignment of forces is preserved. Note that 3D1-68GT demonstrates 20 327 "marks" at 3DMark'03 in the nominal mode. Can you see now why the box shows results just for 3DMark'03? At the nominal frequencies, 3D1-68GT fails to reach the 10 000 bar in 3DMark'05, while at 3DMark'03 it successfully demonstrates 20 000 "marks". The figure is more tempting! :) Going on.

Despite the two "heads" and overclocking, at this resolution the 3D1-68GT can't beat even the single 7800GT. To all appearances, the performance limitation of the CPU does not let 3D1-68GT show its worth. Let's see what will be at higher resolutions.

The assumption proved correct. 3D1-68GT showed results comparable to those for 7800GTX, while 7800GT shows a 10% lag.

Another increase in load upon the video subsystem is another proof of that 3D1-68GT is quite ready for that. While its predecessor 3D1 (one based on two 6600GT) was inferior to "single" 6xxx representatives in the most demanding modes, then 3D1-68GT competes on par in the Supreme League of today.

Low resolution and lack of CPU power smoothes out the difference in results. It seems like 3D1-68GT is very near to "beat" 7800GTX-SLI, but the miracle hasn't worked.

The difference becomes more noticeable. It is no longer possible to encroach upon the supremacy of 7800GTX-SLI, but rivals like 7800GT/GTX are not that close.

The picture does not change radically.

The most demanding game puts all dots on i's. Serious optimizations of video chips for the G70 generation were not lost on them. 3D1-68GT is desperately trying to stay on par. Only overclocking helps.

The situation is a bit better. 3D1-68GT is between two fires. 7800GT is pressing from beneath, and 7800GTX - from above.

The picture does not radically change, but there is one surprise. In the end, overclocked 3D1-68GT failed to pass the whole test throughout. In particular, it did pass it without artifacts, but regularly hung at the last frame of the test, thus we were unable to know the result. At the nominal frequencies, we found no problems. Nevertheless, I left the overclocked 3D1-68GT where it was. The result should be interpreted like this - "here could be the results for overclocked 3D1-68GT, if it were not for the unknown "bug" in the last frame which prevented us from seeing them".

Final Words

No doubt, 3D1-68GT is a very interesting video card. The card is unlikely to gain wide occurrence on the mass market. And, admittedly, it is a bit too late. Certainly, the performance demonstrated by 3D1-68GT is very very good. But as you could see for yourselves on the example of F.E.A.R, it is extremely difficult for 3D1-68GT to compete at new-generation games with "single" leaders of today. Reality is such that with time the difference will be increasingly noticeable. Nevertheless, for now 3D1-68GT is a product that is worth of attention.

The "two-headed" video cards of 3D1 series open another very interesting possibility - in the Internet you can come across messages of successful re-making of NForce 4-4x chipsets into full-featured NForce 4 SLI. Of course, such re-making won't add a full-featured PCI-E slot on the motherboard, but you can install a 3D1 video card into the only PCI-E 16x slot and raise the performance of the video subsystem substantially. You might object - why such perversions are needed? In any case, single 7800GT/GTX show similar performance. I intentionally stress it out - a solution of 3D1 series. The path has been paved, and two-headed solutions based on most recent VPUs will appear sooner or later.

On the other hand, 3D1-68GT has every chance to become a rarity, and who knows maybe in a couple of years you might earn a lot at an auction? : Unfortunately, not all unusual moves to becoming a rarity with time. Would you give a lot of money for the first keyboard that Bill Gates used? Personally, I wouldn't. But his brain encephalogram recorded during writing his first MS-DOS would definitely become a sales hit. Sort of a textbook for beginner programmers. Thing like "strive towards that, and you are in for a great success". Anyway, I'd like to warn the followers of such approach - having decided to write your first program and trying to look as as close as Bill Gates as possible, be careful not to arrive at the same alpha-rhythm, that is, to avoid the same MS-DOS. :) :) Well, I have distracted a bit.

According to the information from Gigabyte Russia's representation office, the recommended price for 3D1-68GT is set to $599. Video cards should arrive within a few weeks. Gigabyte representatives recommend to contact authorized distributors of the company.

Let's sum it all up:
• The requirements to the system imposed for 3D1-68GT are serious enough. Support for SLI is a must, and a powerful PSU, good cooling are also needed.
• The price somehow exceeds the middle level of video cards built on 7800GT/GTX, but is less than the total cost of two separate 6800GT cards, and of course much lower than two 7800GT/GTX.
• The performance of 3D1-68GT at most modern games is close to 6800 Ultra-SLI and surpasses that for single 7800GT/GTX by as high as 25%.
• 3D1-68GT is a stylish image-making thing.

Everyone should decide upon what is the most important of the above listed. No doubt, 3D1-68GT is a successful product albeit belated with the release onto the market.

Conclusion

3D1 based on two 6600GT proved to be the first successful signs for the past two years, although the 128-bit limitation for the GPU but prevented it from showing its worth. 3D1-68GT video card is a dignified heiress of 3D1. I hope engineers at Gigabyte have done a good job mastering the practice of creating video cards on the base of two GPUs and work at next-generation dual-chip video cards is at full sway at laboratories.

Indeed, why not to make 3D1 on two 7800GTX chips at a time? The power consumption of 7800GTX is smaller than in 6800Ultra. 7800GTX video processors support up to 512 MB of memory - a really good chance to be the first in the world to release a video card of 1 GB memory onboard. Although that much is not yet needed, but GIGABYTE will certainly have something to boast unless some rival outdoes it.

Do 3D1 solutions have a right to live? Definitely, yes!

Will they be attractive to the user? The question is a bit more complex to answer. As we already stated, the system requirements for two video cards in the SLI and one dual-chip video card are the same. The system configuration made up of two video cards is more flexible. But dual-chip video cards offer one more potential advantage. Perhaps, eventually NVIDIA will dare for the move like support for four VPUs within one system. Why do you think the company is producing chipsets of the NForce 4 SLI x16 series then? In this regard, installation of two dual-chip video cards of the 3D1 type in a system with two PCI-E 16x slots does not seem so extraordinary.

Will a SLI-bridge be needed to merge two dual-chip video cards? You never know. Today's designs of video cards of the 3D1 series do not provide for that. Perhaps that would require introducing changes to the already complicated wiring. Hopefully, the engineers will overcome all these problems with honor and surprise us with unusual solutions.

No doubt, much will have to be done to let us enjoy the full power of four VPUs fitted within a single system. As rumors have it, the drivers that allow that are already coming up. But there is a substantial restriction from the CPU side - the CPU performance is already not enough sometimes to load even the two video cards. On the other hand, wide occurrence of dual-core CPUs will sooner or later compel the developers to optimize the software in terms of real dual-core performance of CPU and we will see old records renewed.

Finally, another insane fantasy. I wonder is NVIDIA does implement support for 4 VPUs in a single system, will engineers at Gigabyte start creating a "four-headed" video card? What would be its name then? 3D2? 3D4? 3D-QUAD? I think neither engineers at Gigabyte nor marketing people still don't know the answer to these questions. Time will tell.

For now, we are delighted to grant our award "For innovative design" to Gigabyte.