Green Team Roundup - GeForce GTX 200 & 9800 review
Author: Luka Rakamaric
Date: 28 Jul 2008

A couple of weeks ago, NVIDIA launched by far the most complex GPU on the market, the GT200. With its 1.4 billion transistors, it is almost 50% larger than the current AMD/ATi flagship used in HD4850 and HD4870 cards. Today we are going to compare the two cards based on GT200, GTX 280 and GTX 260, but in their SLI configurations. We’ve also thrown in some good old 9800GTX’s, as well as 3-Way SLI GTX 260 configuration, just for flavor..

The GPU

The GT200 GPU is the most complex GPU to date. Intel has one more dual core CPU that has 1.7 billion transistors, but the vast majority of that number is due to L3 cache size of 24 MB, rather than its computing power. GT 200 has roughly two times the transistors of the previous generation flagship, G92. Unlike ATI, NVIDIA didn’t choose to build this GPU in a 55-nanometer production process, so with 2x the transistor count comes 2x the die area. All that would probably mean that this GPU is one big heating machine, but extensive optimizations were introduced to reduce the power consumption to just above the level of 8800Ultra. The GPU now includes 240 stream processors, which is almost twice as much as the G92 and G80. They are now clustered in three groups of eight, with ten such units for a total of 240 (3x8x10). In the previous generation we had a 2x8x8 configuration. Each group of   eight shaders has its own 16k of shared cache, which avoids moving data to the cards framebuffer and back. In the case of GTX 260, there are only 8 units instead of 10 (although there are 10 units physically, 2 of them are disabled), so it only utilizes 192 stream processors.

- GTX 280 : 240 shader cores, 602 MHz core clock, 1296 shader clock, 1107 MHz memory clock, 1 GB frame buffer, 512 bit bus, 141.7 GB/s throughput.
- GTX 260 : 192 shader cores, 576 MHz core clock, 1242 shader clock, 999 MHz memory clock, 896 MB frame buffer, 448 bit bus, 111.9 GB/s throughput.

In addition to disabling some of the cores, there’s also one 64-bit memory controller missing, so instead of 512-bit wide bus, the 260 has only 448-bit interface, which in turn with lower memory clock decreases the memory throughput significantly.

Both cards use GDDR3 memory, while they achieve their high memory throughput  by using a wide bus. In contrast, AMD/ATi’s 4870 uses GDDR5 memory which has four times the effective clock but on a 256-bit bus. The end ratio remains the same, but AMD/ATi’s PCBs are less complex. The downside is that GDDR5 memory is pretty pricey and comes in limited supply.

Both cards are very similar, with the only difference being the power connectors. 280 uses 6+8 while the 260 uses 6+6. Coolers are similar in design to the recent 9800GTX and 8800GTS 512 cards, but more massive, and the cards are shielded from all sides, so there is very little PCB visible. The fan is relatively quiet, but in SLI configurations it can get a little loud, since there’s a limited air supply between the cards. The angled fan enables more air to reach it, unlike the 8800GTX and Ultra designs, but it is still far from ideal.  Both cards have two SLI connectors to enable 3-Way SLI. If you are considering placing three GTX 280 cards in one system, a real brand-name 1 kW power supply is a must (we used PC Power and Cooling Turbo Cool 1KW-SR), and 1200W is what NVIDIA recommends.

For benchmarking SLI performance we used NVIDIA’s latest SLI chipset 790i on ASUS Striker II motherboard. Even though NVIDIA recommends at least 1200W power supply, we used a 1000 watt power supply with a QX9650 CPU with no problems.

We benchmarked the following configurations:
- 9800 GTX
- 9800 GTX SLI
- 9800 GTX 3-Way SLI
- 9800GX2
- 9800GX2 Quad SLI
- GTX 280
- GTX 280 SLI
- GTX 260
- GTX 260 SLI
- GTX 260 3-Way SLI

Unfortunately we were unable to get three GTX 280 cards at publishing time, but scaling wise it shouldn’t be too different from 3-Way SLI configuration of GTX 260. We'll do another article with GTX280 scores very soon.