MultiGPU Update: Does 3-way Make Sense?

Who Scales

As we mentioned, we'll be looking at scaling from both 1 to 3 and 2 to 3 GPUs. This gives us a few more metrics than last time to look at both overall and on a per game basis. From the overall standpoint, we'll first look at scaling from 1 to 3 GPUs. We'll again look at general success as >33.3% scaling and complete failure will be <5% scaling. This will give us information on how many titles seem to be only CPU limited and how many are of zero or negative value as compared to a single card.

Before we get to the numbers, it is important to note that all of this data is out of 21 tests for AMD cards (like the previous article) but out of 20 for NVIDIA hardware. We had an issue with FRAPS running at 2560x1600 with 3-way NVIDIA solutions. We do want to be clear that the game ran fine, and this seems to be a high res high memory usage issue in Race Driver GRID when FRAPS is combined with 3-way and higher SLI. Let's make it clear that this isn't an issue with the game or the hardware per se, but an issue in combination with FRAPS. 3-way SLI runs really well on Race Driver GRID: we just can't tell you how well unless and until this issue is resolved. We leave the game in this article because there is AMD data at 2560x1600 and because there's still usable data at 1680x1050.

First up in our look at who scales is general success (>33.3% scaling) in moving from 1 to 3 GPUs:

The cards that sees the least success in moving from single GPU 3-way are the 4870 512MB and the GTX 285. With the 4870 512MB, this is a combination of failures and CPU/system limited situations while the GTX 285 is purely CPU/system limited here. 16 out of 21 tests isn't hugely different than the 18 or 19 out of 21 (20 for NVIDIA cards), but we do see less "success" in general as compared to our two card situation. And keep in mind that this is 33.3% out of a possible performance improvement of 200%. We are being less strict and seeing less success.

Now lets look at complete failure of scaling from 1 to 3 GPUs. This is based on scaling of <5% and ends up catching the cases of negative scaling. While we did this on a per game basis for 2-way scaling, this time we are looking at the results out of the total number of tests (out of 21 tests for AMD cards and out of 20 tests for NVIDIA)

Again, this is generally failure (negative scaling) at 2560x1600 and is generally an issue we can attribute to 512MB of RAM not being enough at high resolution. There are some differences here as compared to 2-way scaling, but generally this isn't that many cases of abject failure to contend with. We'd still love to see AMD and NVIDIA implement something that caught multiGPU failure and reverted to running on a single card in those cases rather than producing a negative experience. But since we can manually disable both SLI and CrossFire, this isn't a deal breaker (it's just an annoyance).

When we look at scaling up from 2 to 3 cards, things get a bit more dim. Since the maximum scaling percent is 50%, we decided to lower our bar for calling a configuration "successful" by reducing our threshold to 10% (which is very generous at only 1/5th of the theoretical maximum). Our results show that much reduced improvement when moving from two to three cards. Here's the data on the number of "success" (>10% improvement) we saw when going from 2 to 3 cards:

With the best success rate coming in at 14 out of 20 with the GeForce GTX 260 3-way SLI, and our threshold for success so low, 3-way isn't looking so great out of the gate. Let's take a look at failure to round that out. We'll consider failure to be <2.5% scaling.

These numbers show that the majority of the time that cards don't scale up well from 2 to 3 GPUs, they either make no statistical difference or they degrade performance. This is in contrast to scaling from 1 to 2 GPUs. So despite the fact that 3 GPUs can offer good improvement over 1 GPU, it doesn't seem that 3 GPUs consistently offers good improvement over 2 GPUs.

But this is the high level overview. Let's take a look at each game test to get a better idea of what's going on. First we'll recap prices and the test setup and then we'll get to the analysis.