NVIDIA GeForce GTX 590 3GB Dual-GPU Fermi Graphics Card Review

1024 CUDA cores on a card

The “Top Secret” GTX 590 turns out to be both better and worse than the Radeon HD 6990 4GB depending on some vary particular use cases. In realm of $700 graphics cards, this is definitely something you want to pay attention to. But I am getting ahead of myself; let’s first dive into the design on the GTX 590 and see what’s under the hood.

The "Top Secret" GTX 590 turns out to be both better and worse than the Radeon HD 6990 4GB depending on some vary particular use cases. In realm of $700 graphics cards, this is definitely something you want to pay attention to. But I am getting ahead of myself; let’s first dive into the design on the GTX 590 and see what’s under the hood.The High-End Battle Commences

Just a couple of weeks ago AMD released the Radeon HD 6990 4GB card, the first high-end dual-GPU graphics card we have seen released in quite a while it seems.  Before that, the Radeon HD 5970 had been sitting on the throne as the fastest single card for even longer – the GeForce GTX 295 was NVIDIA’s last attempt at the crown.  Even before we got our hands on the HD 6990 though, we were told by various NVIDIA personnel to "wait what we have in store."  Well, we have done so and today we are here to review and discuss NVIDIA’s entry into the dual-GPU realm for 2011, the GeForce GTX 590 3GB.

The "Top Secret" GTX 590 turns out to be both better and worse than the Radeon HD 6990 4GB depending on some vary particular use cases.  In realm of $700 graphics cards, this is definitely something you want to pay attention to.  But I am getting ahead of myself; let’s first dive into the design on the GTX 590 and see what’s under the hood.

Note: If you would like to check out our video comparison between the GeForce GTX 590 and the Radeon HD 6990 before moving on, please do!

Gemini – The NVIDIA GeForce GTX 590 3GB

NVIDIA Gemini boards (the code name for all dual-GPU configurations internally) usually follow a pretty simple pattern: take the highest end available GPU, downclock it some to save on power consumption, slap two of them onto a single PCB and *poof*, instant high-end graphics card.  The same algorithm was followed for the GTX 590 in this case though you will see below that some of those specification changes are worth noting and discussing.

Above you will find the specifications of the GeForce GTX 590 as provided by NVIDIA.  There are 1024 total CUDA / stream processors on the card, 512 on each GPU, equating the GPUs to the GF110-based GTX 580 single GPU solutions we have had around for some time.  However, there is a significant drop in clock speed on these parts.  While the reference GTX 580 cards run at a frequency of 772 MHz, the GTX 590 here is running at only 607 MHz; that is a 27% drop in clock speed per GPU.  Anyone that follows the GPU battles knows that clock speeds can basically report a near-linear performance increase (or decrease) within a margin of error of a few percent.  To put that in perspective, here are some clock speeds of other currently available products from NVIDIA:

• GTX 590: 607 MHz
• GTX 580: 772 MHz
• GTX 570: 732 MHz
• GTX 560 Ti: 822 MHz
• GTX 550 Ti: 900 MHz

Now, we know that because of the different GPU configurations (core counts, etc) that these aren’t directly comparable, but it is interesting to see where the GTX 590 clock speeds stand, and the answer is pretty low on the totem pole.   In comparison, on the Radeon HD 6990 which has a stock clock speed of 830 MHz, we only see a drop of about 6% from the reference Radeon HD 6970 speeds. 

Memory speed sees a similar drop, going from 1002 MHz on the GeForce GTX 580 down to 850 MHz on the GTX 590 – a drop of about 17%.  There is 3GB of total GDDR5 memory, the same 1.5GB per GPU configuration we find on the GTX 580. 

Much like the Radeon HD 6990, the GTX 590 needs to cool a tremendously hot GPU configuration in a single card and thus requires a bit more engineering than the standard offerings.  NVIDIA is using a dual vapor chamber solution to quickly move all ~365 watts power from the surface of the GPU to the heatsink fins to be cooled by the centrally located fan.  The PCB on the card is 12 layers deep and runs with a 10-phase power supply behind it. 

Though these benchmark numbers came straight from NVIDIA, I thought I would provide them to you before seeing our testing so you understand our thought process going in.  Even using NVIDIA’s best case scenario (which they obviously present to the media), the GTX 590 is only ranging from 28-60% faster than the GeForce GTX 580 single-GPU card.  Personally, I found this to be a bit of a letdown and indicative of the severe clock speed drops required to get the GTX 590 up and stable.  NVIDIA’s SLI technology has been pushing the 90% scaling boundary in our testing in recent months so we don’t think there is a huge overhead on the GTX 590 that is software related.

NVIDIA’s testing does show that the GTX 590 is faster than the HD 6990 in this collection of games, though some of our results differ as you’ll see on the following pages. 

They were proud to show that the GTX 590 was significantly quieter than the Radeon HD 6990 – a result we very clearly noticed in our day to day testing.  The GeForce GTX 590 is well engineered for noise levels compared to other dual-GPU offerings.

The output configuration on the GTX 590 is great news for gamers looking to try out NVIDIA Surround or even 3D Vision Surround.  With a set of three dual-link DVI outputs a single mini-DisplayPort connection you can push 4 monitors at once and won’t require more than one adapter to make it happen.

Though not enabled on this initial press driver, NVIDIA is pushing the possibilities of Quad SLI with great scaling potential.  Based on the games selection that NVIDIA went with here (HAWX 1?  AvP?) I wonder if the technology is ready for high-end gaming titles of 2010 or 2011.


Now let’s look at the card itself to see how NVIDIA managed to snuggle a pair of what is arguably the most power hungry of modern GPUs on to a single PCB.