Last month, LucidLogix Technologies announced Hydra, a new take on multi-GPU implementation for desktops and notebooks. Monday, I got a chance to see a live demo of the technology, and get some clarity on what exactly this thing is expected to do.
Before we get into the demo, allow me to provide some context in case the previous link did not do its job (very likely, as looking back on it, it's kinda thin; anyway...). The Hydra 100 is a Silicon on Chip (SoC) solution to scaling 3D graphics. Basically, it allows you to, for example, insert up to four graphics cards from any one vendor (ATI or Nvidia) and receive linear performance from each card. That's the promise at least.
Now you may be asking, "Doesn't this already exist?" Well kinda. Each graphics card vendor has its own solution that allows you to place multiple GPUs into one system to achieve increased performance--ATI with Crossfire and Nvidia with SLI. What separates the Lucid method is the techniques involved in achieving this.
- ATI and Nvidia achieve this in one of four ways in a two-card setup:
- Split frame--Here, each frame (of whatever game you're playing) is divided into two halves, and each half is sent to a different GPU to be processed.
- Alternate frame--Here, each GPU processes the whole frame. While one GPU is processing the current frame, the other GPU starts rendering the next frame, so that when the first frame is rendered, the second frame is ready and waiting.
- Scissors (used only by Crossfire)--It's similar to split frame; however here the technique is more dynamic as it can make the faster card render more than half the screen at a time.
- SuperTiling--Also only used by Crossfire, here the screen is divided into many small squared portions of the screen. Each GPU then is in charge of rendering selected portions. The faster card will be tasked with rendering the most portions.
As good as these techniques work for their respective platforms, they are inefficient, and 100 percent linear performance of each card is not achieved. With Hydra 100, Lucid's goal is to achieve 100 percent linear performance with each card.
While Crossfire and SLI each choose one method and run with it while the game is playing, Hydra dynamically determines which technique will be used on a per-frame basis. Since every game has a different bottleneck and these bottlenecks tend to shift even between frames, Lucid concludes that there is no single solution for all frames and games. Hydra includes a real-time graphics distribution system that determines which method at any given time to use to share the load. "This decision-making is done in real-time on a per-frame basis and introduces no latency," according to Lucid.
The demo I saw used a reference design Hydra-enabled motherboard with two Nvidia 9800GTs installed. During the demo, Lucid showed off how the work is dynamically distributed between the two GPUs. In the pic below, GPU1 is rendering the screen on the left and GPU2 the screen on the right.
Notice how the left screen is not rendering anything that the right one is (except the gun, which was done for display purposes) and vice versa. Also, what each GPU was rendering changed on the fly as they moved through the game.
Lucid promises that you'll be able to mix and match older and newer cards from the same vendor. For example, an Nvidia 9800 will work with an Nvidia 6800. How far back the support goes, however, will probably depend on whether the two cards in question are supported by the same driver. While the ability to mix and match different cards is supported by ATI's Crossfire, it is not something the Nvidia supports with SLI.
Also, just to be clear, you will not be able to use graphics cards from both Nvidia and ATI at the same time. This is because Windows does not support the use of two different graphics drivers while rendering a 3D application.
Lucid is hoping to have the technology available to consumers in 2009, but this will depend on the OEMs they're currently working with. None of which they wanted to reveal to me as of yet.
- Here are the most likely ways you'll see Hydra implemented:
- On a desktop or notebook motherboard, connected to a standard PCIe port between the CPU subsystem and the multiple graphics add-in boards or MXM modules.
- On an add-in board, enabling up to four GPUs by any one vendor, that can be plugged into a single slot on any motherboard.
Lucid says they're shooting for linear performance of each card. So if you have two or three cards from the same vendor, supposedly all will be running at their top speeds with no bottlenecks. Since both Crossfire and SLI do not achieve linear performance for each card, this would be quite a feet indeed. Here's hoping they can make this happen.
If you're interested in a much more technical article on Hydra check out this one here at PC Perspective.