Graphics is infinitely parallel. There is always more work that can be done, and the work can be broken up into millions of completely independent operations. This is very different than most tasks we see done on the CPU that don't scale quite as easily with the number of cores. While we might see small improvements by adding another CPU, we can see nearly double the performance by doubling the number of processors in a graphics card (as long as there are no other bottlenecks anyway). This fact is why AMD and NVIDIA have invested so much money into their respective multiGPU solutions (CrossFire and SLI respectively).

MultiGPU solutions have been around for a few years now, and while we frequently include single card multiGPU solutions in our reviews, we only occasionally take an in depth look at multiGPU technology. Some time has passed since the last time we studied the issue, and now that we've fully broken in our Core i7 system, 64-bit Vista, and recent graphics drivers, it's time to get to it.

Over the past few weeks we've been benchmarking and analyzing lots of numbers. We've looked at single, two, three and four GPU systems across multiple games and resolutions. The configurations we chose to look at are current generation high-ish end hardware capable of operation in 3-way and 4-way configurations. Because of the sheer volume of data we collected, we've decided to break up our analysis into multiple articles.

This first article (the one you're reading right now) will cover single and dual GPU configurations (including single card multiGPU hardware). The next article will add 3-way solutions along with comparisons back to single and dual GPU setups. The final article will add in 4-way performance analysis and compare it back to the single, dual and 3-way data. Splitting up the analysis this way will allow us to dive deep into each type of configuration individually without spreading the content too thin. We can keep focus on a specific aspect of multiGPU performance and scaling while still making all the relevant comparisons.

The initial installment also introduces the Sapphire Radeon HD 4850 X2 2GB. Though we expected AMD to push the 4850 X2 out in the same way they launched the 4870 X2, we've only seen one version of the 4850 X2 hit the scenes late last year from Sapphire. In light of what we've seen, we are rather surprised that we haven't seen more fanfare behind this part from either AMD or other board makers. The lighter weight X2 competes more directly in price and performance to the GeForce GTX 280/285, and really fills out the lineup for AMD. Overall, the increased RAM in the 4850 X2 2GB enables great performance scaling even at resolutions the 512MB 4850 can't come close to handling.

As for the topics we'll cover, our interest will focus on scalability of the multiGPU solutions and the relative value of the same. Before jumping into the numbers, we'll cover the metrics we use to analyze our data. First, we'll look at scaling and talk about the big picture. Then we'll talk about what we've done to calculate a value comparison.

Who Scales: How Often?
POST A COMMENT

95 Comments

View All Comments

  • kmmatney - Monday, February 23, 2009 - link

    Especially at the 1920 x 1200 resolution - that resolution is becoming a sweetspot nowadays. Reply
  • just4U - Monday, February 23, 2009 - link

    I disagree. I see people finally moving away from their older 17-19" flat panels directly into 22" wide screens. 24" and 1920/1200 resolutions are no where near the norm. Reply
  • SiliconDoc - Wednesday, March 18, 2009 - link

    Correct, but he said sweet spot because his/her wallet is just getting bulgy enough to comtenplate a movement in that direction... so - even he/she is sadly stuck at "the end user resolution"...
    lol
    Yes, oh well. I'm sure everyone is driving a Mazerati until you open their garage door....or golly that "EVO" just disappeared... must have been stolen.
    Reply
  • DerekWilson - Monday, February 23, 2009 - link

    The 1GB version should perform very similarly to the two 4850 cards in CrossFire.

    The short answer is that the 1GB version won't have what it takes for 2560x1600 but it might work out well for lower resolutions.

    We don't have a 1GB version, so we can't get more specific than that, though this is enough data to make a purchasing decision -- just look at the 4850 CrossFire option and take into consideration the cheaper price on the 1GB X2.
    Reply
  • politbureau - Tuesday, June 1, 2010 - link

    I realize this is an older article, however I always find it interesting to read when upgrading cards.

    While I find it admirable that Derek has compared the 'older' GTX 280 SLI scaling, it is unfortunate that he hasn't pointed out that it should perform identically to the GTX 285s if the clocks were the same.

    This was also passed over in the "worthy successor" article, where it does not compare clock for clock numbers - an obvious test, if we want to discover the full value of the die shrink.

    I recently 'upgraded' to 3 GTX 285s from 3 GTX 280s through warranty program with the mfg, and there is little to no difference in performance between the 2 setups. While cabling is more convenient (no 6 to 8 pin adapters), the 285s won't clock any better than my 280s would, Vantage scores are within a couple hundred points of each other at the same clocks (the 280s actually leading), and the temperature and fan speed of the new cards hasn't improved.

    I think this is a valuable point in an article that compares performance per dollar, and while slightly outside the scope of the article, I think it's a probabtive observation to make.
    Reply

Log in

Don't have an account? Sign up now