eVGA e-GeForce 7800GT CO 256MB

Performance:

Although we are not delivering 7800GT SLI results in this review, we used the Tyan Tomcat K8E-SLI S2866 motherboard that is presently our top performing motherboard when it comes to a Socket 939 solution with SLI abilities and is based off the nForce Professional 2200 Chipset rather than the nForce4 SLI. Below are the hardware and software components used throughout the duration of testing.

In addition to testing the eVGA e-GeForce 7800GT CO, we also ran the eVGA e-GeForce 6800GT 256MB, Leadtek PX7800GTX 256MB, and a pair of GeForce 6600GT 128MB in SLI through its paces. Upon installing the 7800GT and the 1.0-8174 x86_64 NVIDIA drivers, we launched nvidia-settings and were enlightened by its detection. For further reference we compiled the latest version of NVClock (0.8 BETA) and the 7800GT was found. The architecture was detected as being NV47 A1 rather than G70. Additionally, the onboard internal sensor was detected but between nvidia-settings and NVClock, there was a 2°C discrepancy in the temperature but even so, the unit operated well within our temperature threshold. Speaking with Roderick Colenbrander, the author of NVClock, v0.8 BETA uses a fixed offset in its temperature correction but studying the source for a RivaTuner v15.8 plug-in he has created a fix for the temperature display and the code will be added to the CVS tree. In addition, the GPU clock speed of 470MHz was recognized as well as the 1100MHz 256-bit memory frequency. One of the things we recognized thanks in part to Roderick's program, is that although there are three voltage levels with the card of 1.20, 1.30, and 1.40V, what remains to be seen is any additional performance levels for the card. The BIOS is simply set to run at 470/1100MHz with 1.40V and there is no set scaling for 2D frequencies or alternative speeds. We imagine this was set in part due to eVGA factory overclocking the card from its BIOS and really is not a factor except for a small amount of additional heat generated when in 2D mode as well as the fan speed is always maxed out. Although the 7800GT CO remained thermally cool in all of our testing, we noticed it was a bit of a screamer compared against its competition including the 7800GTX 256MB. However, the system we used in testing was open-air and those running their system inside of a chassis should not notice much of an impact.

Starting off from 470/1100MHz (GPU/MEM), which is 70MHz faster than NVIDIA's standard G70 GT core and 100MHz faster on the side of memory, we set out our overclocking expedition with CoolBits enabled through nvidia-settings as well as using NVClock 0.8 BETA. We were surprised to find that we were unable to push the GPU any faster than 470MHz; even specifying 471MHz would be recalled and would revert to the hardware defaults. We believe the GPU problem is due to eVGA's video BIOS and a conflict with the Linux overclocking tools available. Even so, the 470MHz frequency is already vicious for the GT part and thanks to this, it should be able to compete with the more expensive 7800GTX. Moving onto the memory overclocking, we were most impressed by how far we were able to push the GDDR3 video memory. Starting from 1100MHz, we had no problems pushing the memory up to a maximum value of 1260MHz, a 160MHz overclock. While running at 470/1260MHz we faced no artifacts or stability sacrifices while doing simple tasks inside GNOME, but once we had turned to run some benchmark trials we faced interim faults where we had noticed the occasional artifact and ultimately faced stability challenges. Backing off to 470/1230MHz, we were liberated of any artifact or distortion issues and the NVIDIA card was able to pass 100% of our tests. Compared to the reference speeds, we were most pleased with a 70MHz stock (17.5%) and 230MHz (23.0%) overclock.

Recently we have revised our graphics card benchmarking investigation to provide more thorough and additional numbers than what can be found in our numerous previews, primers, and display driver comparisons. The benchmarks we are running on the graphics card are Enemy Territory, Doom 3, Quake 4, Unreal Tournament 2004, and SPECViewPerf with all of them being real-world gaming benchmarks except for the last benchmark. Standard Performance Evaluation Corporation and its SPECViewPerf product offer numerous view-sets from 3D Studio Max to Maya to SolidWorks and Pro/ENGINEER in order to represent OpenGL workstation performance. For those who have attempted to use UnrealMark Linux BETA 3 on an x86_64 system and faced a segmentation fault when booting the application, the problem can most easily be resolved by compiling UMark on a similar 32-bit system and transferring over the resulting binaries. During Unreal Tournament 2004 testing we used the demo version and used 12-bots as the basis for the testing. As usual, our various Phoronix practices and settings applied during the complete benchmarking and testing phase. With the GeForce 6600GT pair being the only set of graphics cards in this article running NVIDIA SLI, we ran the multi-GPU technology in Alternate Frame Rendering (AFR) mode.