   |
|||
|
|||
|
:
News
: Archives
|
|
Gigabyte 8600 GTS Silent Pipe 3 |
|
|
Quiet Performance |
|||
|
|
|||
|
by Josh Walrath |
|||
|
NVIDIA’s release of the 8600 series was a long awaited event. Last year NVIDIA unveiled the 7600 series, which became a favorite for the price/performance crowd. The 7600 performed on a level that was higher than the previous 6800 GT and 6800 Ultra, and was being offered at price points below $239. People expected this to be the next “GeForce 6600 GT”, and they were right. The card has sold in droves, all the while decreasing in price over time. The 7600 GT is still a popular card, namely because it has dropped to around $110 US. The big question now is if the 8600 series, and namely the 8600 GTS, can successfully carry on the 6600/7600 tradition. Today’s review deals with Gigabyte’s entry into the 8600 GTS world. This one is actually quite special, as it relies on a very elaborate cooling system. And it has no fan. Gigabyte is one of the big 3 Taiwanese motherboard manufacturers, and as such they have branched into the graphics card world. Mostly their designs were based on reference boards, but in the past year they have been trying to carve out their own niche. Some of their midrange products have sported non-reference designs, as well as other fanless cooling solutions. I can certainly appreciate what Gigabyte is trying to do, as the “cookie cutter” class of video cards provided by a multitude of vendors is a rough market to be in. The G84 Chip Last November NVIDIA released their first G8x product, the 8800 GTX and GTS. The G80 chips these models are based on are simply massive. 680+ million transistors and a die size of over 480 mm square made this a monster. The performance and features of the chip were also monstrous. The challenge for NVIDIA was to be able to cut down the G80 into more manageable portions for the midrange and budget sectors. The 8600 series is based on the G84 chip, which is the result of the cut-down G80 project. The first difference between the G80 and G84 is that the G84 is made on TSMC’s 80 nm process, while the G80 was on the older 90 nm node. Using the smaller node made the G84’s die size approximately 169 mm square. For reference, the older GeForce 7900 GTX used the G71 chip, which was comprised of approximately 289 million transistors and was 196 mm square. This is not exactly an apples to apples comparison, as the G84 is a significantly different design than the G71. The G84 also features approximately 289 million transistors, but a greater portion of these are dedicated to onboard caches unlike the older G71. The G84 has a total of 32 stream processors, 8 ROPS, and 16 texture address and filtering units. Because the stream processors are scalar units, they only work on one component at a time. When running applications which utilize, on average, a Vec3 + Scalar instruction, the G84 acts more like a traditional 8 shader unit. The 32 stream processors do overcome this potential limitation for two reasons. The first is that the speed of the units can be as high as 1.45 GHz, so its potential output can be compared to another chip running 16 traditional shaders at 725 MHz. The second is that the scalar units are kept busy all of the time. Traditional shaders were not fully utilized unless they were given Vec3 + Scalar instructions, so if a unit was given a Vec1 instruction then it was being underutilized. With the scalar processor design of the G8x series, they approach 100% efficiency (assuming that threads are not identical between the different units). The unified architecture also allows the chip to dynamically allocate resources between vertex and pixel shading operations. So when an application requires more vertex work, the stream processors are allocated to those duties. With DX10 we see the introduction of geometry shaders as well.
The box inside the box is sturdy, but a little colorless. The chip also shares all of the features of its bigger brothers, the 8800 series. This means that image quality is identical; the only difference is overall performance. The G84 does bring a few more things to the table. The texture address units can handle twice the work per clock, which improves overall texturing performance. The second major improvement is Pure Video 2. This enhanced video playback unit offloads more work from the CPU when playing back H.264 and VC1 content as compared to the older Pure Video unit. This will be quite important when Blu-ray and HDDVD content become more widely available on computers and laptops. In terms of visual quality the G84 encompasses all of the same features of the 8800 series. This means that the new Coverage Sample AA modes are available, gamma correct AA, and my personal favorite, high quality anisotropic texture filtering. Visually it is hard to get much better than this. Plus the minimal hit of going from 4X AA to 16X AA is well worth the quality increase. NVIDIA certainly made some interesting choices with this product, and there are those that already look askew at a few of these decisions. The first issue many have is why there are only 32 scalar units instead of the expected 64? Also, why use only a 128 bit memory bus? I think a good portion of this is architectural balance, as well as simple economics. I can honestly say though, I would have been much happier if NVIDIA had followed their own example with the 8800 GTX and GTS and used non-standard memory path widths. If NVIDIA had used a 192 bit memory bus instead of 128 bit, they would have allowed slightly better bandwidth while using cheaper memory chips. For example, by only using 800 MHz GDDR3 parts, they could have achieved an impressive 38.4 GB/sec of bandwidth. It would have also allowed for inexpensive 384 MB cards. Using the more expensive 1 GHz GDDR3 would give an impressive 48 GB/sec of bandwidth to their midrange product.
If you have found this article interesting or a great help, please donate to this site.
Copyright 1999-2007 PenStar Systems, LLC. |
|||
