Conclusions

Our tests indicate that NVIDIA has a good reason to cut prices of 9800 GTX (and GTX 260) and launch overclocked 9800 GTX+.

This is another sign that AMD has launched a very successful product. While RADEON HD 4870 occupies the price segment below 300 USD (real prices will be higher at first, as it always happens with new products), but its performance sometimes reaches the level of the more expensive GTX 260! Now we understand why NVIDIA has started cutting GTX 260/280 prices, despite lower profits. Such cards are very expensive to produce due to 512-bit layout (even though it's actually 448-bit), expensive GT2xx GPU, etc.). So, we can congratulate AMD with the long-awaited successful debut.

Speaking of the mass-produced card from HIS, it's a copy of the reference card. Though I'm grateful for the bundled screwdriver, it always comes in handy. :)

We have also analyzed the effect of memory bandwidth on performance. The charts include results of the HD 4870 operating at 625 MHz (like HD 4850), memory frequency being the same 900 (3600) MHz. The results were quite surprising. In most cases performance didn't depend on memory bandwidth, being limited by GPU. Thus, we retract words about HD 4850 being cut down much in terms of memory bandwidth, and that increasing memory frequency would strongly benefit its performance. Turns out we were wrong. As HD 4850 has a much more powerful core than HD 3870, and its memory bandwidth is even a tad lower (1000 MHz vs. 1125 MHz), it turns out that HD 3870 has excessive memory bandwidth. Or, perhaps, AMD miraculously optimized RV770 (compared to RV670), so that higher performance did not result in higher memory bandwidth requirements.

If GeForce GTX 2xx gets cheaper by a hundred dollars or more, the battle will become really fierce.

The Excellent Package award goes to HIS RADEON HD 4870 512MB.

For more information about performance of various cards, please visit our monthly special i3DSpeed.

Afterword

A few words about our benchmarks.

In our updated article about FRAPS, we have illustrated how crude and inaccurate tests with this utility are. Testers do not have other tools, except for benchmarks built into games.

This article and its first part explain that it's sometimes possible to test games with integrated and identically looped demos. Although this implies lots of potential errors, since measurement accuracy depends on tester: whether he starts/stops FRAPS in time or does it too late/early.

But I have run across situations, when demo load changes abruptly in the very beginning or end. So, a half-second delay in starting/stopping the utility changes the average FPS by 15-20%. That's not a measurement error anymore - such test is a total waste of time. One time you delay the test, another time you start it too early (not intentionally, of course), and you end up with absolutely different performance results.

But even that's not the most important thing. The fact is, there are almost no games with built-in demos anymore. So, testers are forced to use a method that we deem totally unacceptable. They measure gaming performance by walking a straight line from the a starting point in a scene to a selected destination (the nearest fence, tree, etc.).

We all understand that it's impossible to navigate to a finish spot in precisely the same route with different cards and in different resolutions. Besides, such games always introduce random elements into a scene, and objects may be placed slightly differently on the same scene.

Unfortunately, websites that publish a huge number of tests do not always reveal their test methods for each game (except for those with built-in benchmarks).

So, we believe that it's better to offer a limited number of game tests, but each will be crystal clear, accurate and showing actual differences between graphics cards.

Write a comment below. No registration needed!