The Test and Performance Improvement

Here's our test setup:

Test Setup
CPU Intel Core 2 Extreme QX9770 @ 3.20GHz
Motherboard EVGA nForce 790i SLI
Video Cards ATI Radeon HD 4870 X2
ATI Radeon HD 4870
ATI Radeon HD 4870 1GB
NVIDIA GeForce GTX 280
NVIDIA GeForce GTX 260
NVIDIA GeForce GTX 260 Core 216
Video Drivers Catalyst 8.7
ForceWare 177.34
Hard Drive Seagate 7200.9 120GB 8MB 7200RPM
RAM 4 x 1GB Corsair DDR3-1333 7-7-7-20
Operating System Windows Vista Ultimate 64-bit SP1
PSU PC Power & Cooling Turbo Cool 1200W

Performance Improvement

So we'll spoil it right up front and give away the answer to the question on everyone's mind: how does performance improve with the addition of the extra 512MB of RAM to the 4870? And since we got a nice little surprise last week with the GTX 260 Core 216, we can even compare performance improvement of the slightly upgraded models of both the NVIDIA and AMD parts.

These comparisons are taken from the highest playable resolution in each game we tested, which is either 1920x1200 or 2560x1600 depending on the game. This class of card is able to handle the highest resolutions, but sometimes not with all the features cranked up (as our tests are designed). These high resolutions do the best job at stressing both memory and processing power, as processing more pixels every frame has an impact on both.

Another factor to consider is that we can't really tell you the maximum potential theoretical performance gain from adding more memory to a system. With the NVIDIA GTX 260 core 216, we know the maximum theoretical improvement is something like 12.5 percent. This is because we added 12.5 more compute resources. But doubling the amount of RAM, we aren't really doing anything directly to performance: we're just increasing the availability to resources to the hardware which may or may not improve utilization. Adding more RAM decreases the chance that something will need to be pulled in from system memory.

Both routes have the potential to improve performance, but both also speak to the balance of the initial design. We really don't want more memory on a board than we need to adequately feed the GPU, and, at the same time, we don't want so many compute/texture resources on the GPU that we can't possibly feed it enough data to crunch. From our perspective, it looks like the Radeon 4870 with 512MB was targeted at 1920x1200 or lower resolutions. More RAM has a lessened impact on lower resolutions, and the price point of the 4870 is generally in line with what a gamer would love to have paired with a high end 1920x1200 panel. People that buy 30" panels for gaming are more likely to go after more expensive solutions.

Some games also show a benefit from more RAM regardless of resolution, meaning they are very resource intensive games. While you can't make purchasing decisions based on future-proofing (there's really no way to accurately predict what card will do better in the future), the trend has generally been that newer games use larger and more textures and pile on more effects which take up more space in local memory. While the games that benefit across the board now are few, it is possible we could see that number increase down the line.



Oblivion and Assassin's Creed are the only two tests we see that don't see better improvement on AMD hardware. Perhaps not-so-coincidently, Oblivion is also one of two tests we ran where the NVIDIA GeForce GTX 260 Core 216 leads the AMD Radeon HD 4870 1GB (and the only test we ran where the original GTX 260 leads the 4870 1GB) - the other being The Witcher. This does change things up again, but it also highlights that the GTX 260 and 4870 are fairly well matched in general. With minor tweaks to performance they we are seeing a back and forth on whose part leads in our benchmark suite. It's like we have NVIDIA and AMD playing that game where each one grabs slightly higher on a stick.

Unlike the core 216, we've known this 1GB 4870 was coming down the pipe for a long time, and we've honestly expected some performance gain at high resolution. But we really didn't expect this much of a difference. The differentiation between the products is better with 4870 1GB than on NVIDIA hardware.

Let's drill down and look at individual game performance to get the rest of the picture.

Index Age of Conan Performance
Comments Locked


View All Comments

  • JarredWalton - Thursday, September 25, 2008 - link

    I would hope that someday soon AMD will address this with drivers or something... but seriously they dropped the ball here. I mean, 4850 and 4870 are the same GPU, so the only difference is clock speed and voltages. You can't expect me to believe that in this day and age they can't get clock and voltage adjustments to work on-the-fly. A BIOS flash can work it seems, but that just begs the question: why wasn't the BIOS programmed "properly" in the first place? (Possibly they discovered in testing that there were problems with the different voltages?) Users should *NOT* have to flash a GPU BIOS for stuff like proper power saving.
  • Finally - Thursday, September 25, 2008 - link

    Hmm. You are the test-guy. You should know (and tell us, please! :p)

    @HD4870vs.HD4850: You forgot 1 thing: the HD4870 has GDDR5, but the HD4850 has GDDR3. As it has been proven, this makes a big difference. So you can't say that they are the same. GDDR5 seems to be much more undervolting and power-saving-friendly.
  • Spoelie - Thursday, September 25, 2008 - link

    Could you check with ATi if powerplay (down throttling clockspeed *and* voltage) is in the pipeline for a future driver release?

    I've been hearing forum voices saying "it's in the next release" for quite some time now.

  • Jedi2155 - Thursday, September 25, 2008 - link

    In higher resolutions, I think this is a reality.

    I think the article is very truthful and quite a few other sites and come to back this up.

    There is a problem with the frame buffer at higher resolutions and settings, especially if you understand how anti-aliasing among other things work.

    Use Rivatuner to check the memory usage on the frame buffer yourself at those resolutions....
  • NullSubroutine - Thursday, September 25, 2008 - link

    I don't have a problem with them using 8.7 for the 4870, far as I have last heard its a great driver for that card. But that is a horrible driver to use for the 4870 X2. While it wasn't the card being looked at, it can skew the results if you are trying to decide which card to get.
  • Tiamat - Thursday, September 25, 2008 - link

    Page 2:

    512MB -> 1024MB is a 100% improvement (i.e. double the ram) not 50% improvement. 50% improvement would have been to 768MB ram.
  • DerekWilson - Thursday, September 25, 2008 - link

    heh ... you are quite right. sorry about that. i'll fix this.
  • Spoelie - Thursday, September 25, 2008 - link

    Power consumption: "Significantly" more in *both* idle and load?

    idle yes, load no
  • DerekWilson - Thursday, September 25, 2008 - link

    by significant i mean the differences is not negligible
  • Diosjenin - Thursday, September 25, 2008 - link

    I was rather under the impression that the 1GB per/2GB total RAM on the 4870 X2 was generally the reason it could be found in many cases to scale better than two 4870s, since the latter option included only 512MB per/1GB total RAM.

    Now that we have 4870s with 1GB RAM, can you stick two of them together and do a 2x 4870 1GB vs 4870 X2 comparison to see how that can affect the scaling disparities we've seen there before?

Log in

Don't have an account? Sign up now