Summary: Last Wednesday, NVIDIA unveiled their latest high-end desktop GPU, the GeForce 7800 GTX, establishing new levels of performance in the 3D market. At the time we were uncertain how well cards based on its G70 graphics core would scale to higher clock speeds, as its 0.11-micron manufacturing process was designed for value, and had never been used on such a large chip before. In today's article, Brandon overclocks three different GeForce 7800 GTX cards, including one retail board from EVGA, to see how well the 7800 GTX overclocks. Is the chip a dog when it comes to overclocking, or is G70 a real screamer? Find out in this article!
Part of this is made possible thanks to the board’s new pipelines, up from 16 pixel pipelines in GeForce 6800 Ultra to 24 in the GeForce 7800 GTX, while on the vertex processing side, the GeForce 7800 GTX sports eight vertex units, two more than GeForce 6800 Ultra, which contained six units. NVIDIA has also incorporated improvements in each of these units to speed up processing time, as well as improved the chip’s ability to handle floating-point math, which is used increasingly more often in today’s shader-heavy games. NVIDIA has also improved the 7800 GTX’s ability to fetch textures.
On the manufacturing side, NVIDIA played it safe with GeForce 7800 GTX, relying on TSMC’s 0.11-micron manufacturing process. As we mentioned in our 7800 GTX preview article, 0.11-micron is a die shrunk version of TSMC’s 0.13-micron process (only without low-k dielectric), and is considered TSMC’s value process, designed for low costs, rather than high clock speeds. This is why both ATI and NVIDIA chose 0.11-micron for their mainstream and value products, the GeForce 6600/6200 family, and the RADEON X300/X600/X700 series. The cost savings afforded by 0.11-micron played a critical role in allowing ATI and NVIDIA to integrate more pipelines into the GeForce 6600 and RADEON X700 in particular.
NVIDIA decided to use this process again for the 7800 GTX, rather than take a risk on TSMC’s newer, smaller 0.09-micron process, which is intended for higher clocks. NVIDIA got burned during the 0.15-micron to 0.13-micron transition with GeForce FX (from GeForce4), with the part never yielding well at the clock speeds intended. NVIDIA likely didn’t want to take a chance again, opting to play it safe instead, even though that limited them to lower clock speeds. NVIDIA’s engineers must have felt that the shader optimizations they’d implemented, plus the additional pipelines would be enough to get them by, and right now that strategy appears to be the right one, as there are rumors that the launch of ATI’s next-gen chip, R520, has been delayed due to poor yields.
The only downside to 0.11-micron is as a value process, it doesn’t contain the performance optimizations found at 0.13-micron and 0.09-micron that allow for high clock speeds. You can see this today on cards built on 0.11-micron, such as the X800 XL. Our overclocking results with graphics cards based on this chip in the past have been rather mundane, with none of the six cards we tested in our RADEON X800 XL Roundup overclocking to more than 10% of their stock clock speed (one of the cards barely overclocked 5%). On the NVIDIA side, we had a little more luck with the GeForce 6600 GT cards used in our 6600 GT AGP Roundup, but in that case, we’re dealing with a much smaller 8-pipeline chip, with only 146 million transistors. GeForce 7800 GTX is much bigger than 6600 GT, and contains twice as many transistors at 302 million.
With this in mind, we were worried with how well the chip would scale to higher clocks. That’s what we’re here today to find out!
To get an indication of how well the GeForce 7800 GTX scales, we decided to overclock our reference boards sent to us by NVIDIA, as well as EVGA’s e-GeForce 7800 GTX card. EVGA’s board differs slightly from the NVIDIA reference board, with a core clock speed of 450MHz, 20MHz higher than the stock 7800 GTX speed of 430MHz. The memory speed remains the same as the 7800 GTX reference board, at 600MHz.
Other graphics manufacturers are doing this with their 7800 GTX boards as well, suggesting that the chip could have a little bit of headroom for overclocking...
How to overclock your GeForce 7800 GTX
For years, NVIDIA’s Coolbits registry modification has been our favorite method of overclocking GeForce cards. NVIDIA has quietly provided this technique to any NVIDIA card owner in their driver, although it’s officially unsupported.
To overclock your board, simply open the Windows Registry Editor by typing "regedit" in the Run Menu of Windows. Then open the following directory:
"HKEY_LOCAL_MACHINE\\Software\\NVIDIA Corporation\\Global\\NVTweak". Next, right-click on the NVTweak key and create a new dword value named "Coolbits." Adjust the value of Coolbits to "3" by double clicking on it and entering "3" in the value data field. Once you’re done, restart Windows to apply your new settings.
After a quick Windows restart, you’ll now see a new field in your driver’s control panel titled “Clock Frequency Settings”, this is where you’ll overclock your 7800 GTX card (or any other NVIDIA card for that matter). Select “Manual overclocking” if you’d like to adjust the graphics core and memory speeds yourself, or “Auto overclocking” if you’d like the graphics driver to determine the highest speeds based on temperature (ala ATI OVERDRIVE. In SLI mode, the manual overclocking mode is your only option available.
So how far were we able to overclock our GeForce 7800 GTX boards? Pretty far actually. Our best overclock was achieved by one of the NVIDIA reference boards, which hit 505MHz core/670MHz memory (1.34GHz effective).
This is an improvement of 15% on the graphics core, and 10% on the memory.
The second reference card topped out at 489MHz core/658MHz memory, still not bad at all in our opinion. Our EVGA card also maxed out at 489MHz core, with the memory hitting 638MHz.
Of course, we couldn’t stop at just testing a single-card GeForce 7800 GTX configuration, we were also curious to see how two overclocked 7800 GTX boards fared when running in SLI. The highest clock speeds we could attain without artifacts and complete stability was 476MHz core/640MHz memory. This was achieved using the two NVIDIA reference boards (in all honesty, the EVGA is built on the same reference board design, only it ships with higher stock clocks).
The clock speeds are a little bit lower because the temps of the SLI configuration are higher. The master 7800 GTX board located in the first PCI Express slot will always operate hotter than the slave card below it, as the card’s fan is obstructed by the slave board, cutting off its air supply. In NVIDIA’s upcoming ForceWare 80 series driver, end user’s will be able to overclock the master and slave cards independently of each other, which will allow you to run your slave board a little bit faster than the master board for better performance.
To measure the effect overclocking has on temperature, we used NVIDIA’s built-in hardware monitoring feature to watch idle and load temperatures. The load temps were recorded following a looped 3DMark 05 session.
As you can see, overclocking the SLI configuration has a profound effect on the cards, particularly under load. Idle temps went up by five degrees Celsius, from 49 degrees to 54. Under load, the delta was more pronounced, as stock temps went from 72 degrees to 80 once the master board was overclocked.
Overclocking lone 7800 GTX card bumped up the idle temp from 43 degrees Celsius stock to 50 degrees, while load temp peaked at 71 degrees.
Pacific Fighters (kamikaze demo)
3DMark 05 – Direct3D
3DMark 05 – Direct3D
Pacific Fighters - OpenGL
Far Cry – Direct3D
IL-2: FB – OpenGL
LOMAC – Direct3D
DOOM 3 – OpenGL
Half-Life 2 – Direct3D
Half-Life 2 – Direct3D
Splinter Cell – Direct3D
Battlefield 2 – Direct3D
Apparently, these concerns were incorrect, as the three 7800 GTX boards we currently have all seem to scale much better than we expected -- all of them hit clock speeds we’ve never come close to touching with the “less complex” X800 XL.
Based on this, it’s no surprise to see so many 7800 GTX board partners overclocking their boards from the factory. It certainly appears as if NVIDIA played it pretty conservative with the 7800 GTX’s clocks.
We saw some pretty tangible performance gains from overclocking, Battlefield 2 performance went up from 63.7 frames per second (FPS) at 1600x1200x32 with 4xAA/16xAF to 74.7 FPS with the overclocked 7800 GTX board. This is a performance improvement of 15%!
Of course, we must also admit that the gains were pretty slim when running with two 7800 GTX cards running in SLI. This is because with two GeForce 7800 GTX cards running in tandem, our Athlon 64 FX-55 processor is holding the 7800 GTX cards back, we’re CPU-bound in many cases at resolutions as high as 1280x1024! In addition, our tests with Lock On: Modern Air Combat, IL-2 Sturmovik: Forgotten Battles, and Pacific Fighters all show that we’re a little CPU-bound with the 7800 GTX. Because of this, flight sim enthusiasts may want to hold off on SLI entirely unless they plan on gaming at 2048x1536.
In closing, we’re glad to see that the GeForce 7800 GTX scales so well. This must come as very reassuring news to the overclocking crowd. SLI users likely can't wait to dabble with independent overclocking.
Now we can’t wait to see what the community does with the 7800 GTX. We have a strong feeling it will only be a matter of time before someone manages to mount a water cooler to their board. Let the overclocking begin!
|© Copyright 2003 FS Media, Inc.|