Forget Expensive Video Cards

Anonymous Reader writes "Apparently, the $200 in video cards does not produce the difference. While $500 video cards steal the spotlight on review sites and offer the best performance possible for a single gpu, most enthusiasts find the $300 range to be a good balance between price and performance. Today TechArray took a look at the ATI x1900xtx and Nvidia 7900gtx along with the ATI x1800xt and Nvidia 7900gt."

Comments

Re:Not directly related to TFA

[TheRaven64]

Render farms don't need or have video on each processor ... you just have thousands of machines in racks processing data streams.

Actually, that's not quite true these days. A modern render farm has a GPU (or two) in each node, and uses it for all sorts of things. If you are only doing relatively low-quality renderings, you can use something like Chromium and get high framerate, enormous images rendered through OpenGL. If you are doing ray tracing, you can speed this up hugely using the GPU.

Even volume rendering runs on the GPU these days. You can split an enormous volume into 256^3 cubes, render these quickly on an large array of GPUs and then composite the individual rays using the alpha blending hardware on a smaller array of machines in a tree configuration until you have the final image[1].

So, no, not every node needs a video output capability, but if you want state-of-the-art performance they do all need at least one GPU.

[1] Some people are using other kinds of stream processor for this step these days, but that's still a relatively young research area.

Re:What am I missing out on ?

[TheRaven64]

I read a paper from Microsoft Research[1] about rendering text using the GPU. The idea was that the raw bezier paths of each character in a font would be loaded onto the GPU and then each character would be created on the fly by a shader. This gave a huge performance benefit; it reduced the GPU-RAM bandwidth requirement hugely and allowed the CPU to offload pretty much all text rendering to the GPU.

For comparison, take a look at Apple's Quartz 2D Extreme. This uses the CPU to render each character to a texture and stores them in the graphics RAM. These are then composited by the GPU. The downside of this, of course, is that the CPU needs to render the text for every size at which it is used. Even so, this gives about an order of magnitude better performance than the traditional way (and, of course, lower CPU usage).

If this becomes mainstream then a GPU with fast shader support will give:

• Faster text-rendering performance.
• Lower CPU usage when rendering large amounts of text.
• The ability to have effects like Apple's Exposé, but with sharp, fully anti-aliased text at all stages in the zoom effect (the performance on current generation GPUs was fast enough to render entire screens full of small text every frame).

[1] See? They do actually do interesting things. It's a real shame nothing from MS Research ever seems to make it into shipping products though.

Re:Not directly related to TFA

[TheRaven64]

Do you know which cards are commonly used? Are they $500 gaming cards? Cheaper gaming cards? More specialized cards?

As with everything else in a cluster, it's usually whichever has the best price:performance ratio. I'm more familiar with the ones that exist in academia, and these tend to be 'whatever the fastest that we could afford when the cluster was built.' An average cluster node costs around £2000 and upwards. They usually have at least two CPUs, a couple of GBs of RAM (minimum). The less cheap ones will have a high-speed interconnect, adding £500-£1000 to the price of a node (plus more expensive switches), while the cheap ones will just use gigabit ethernet. Adding a £200 GPU adds 5-10% to the cost of the node, while giving up to around a 500% performance increase in many tasks.

Usually they don't need access to the driver code. On *NIX (excluding IRIX) they tend to just run an X server on a display that's not connected to anything and run shader programs on it. The limitation of this is that only one program/user can typically access the GPU at once, but that's usually what's wanted. The shader program receives data from the interconnect, processes it, and passes it on.