Printing-ink veterans don't take cyberspace journalists too seriously - Roy Greenslade, Guardian Online
IF YOU THOUGHT today was only going to be about AMD CPUs, the company has a streaming surprise for you. Up until now, it was pretty quiet about the strategy, but by this time next month, Stream will be everywhere.
The big bang is that Stream consists of three major parts, the GPU, the Compute Abstraction Layer (CAL), and tools. You know about the GPU, and tools are just that, tools for writing code, not really end user focused. The big one is the CAL.
Normally, with something as esoteric as GPGPU, you need to download drivers, patches and software, followed by an afternoon of tuning. It puts the technology in the proverbial geek ghetto, not something for the faint of heart, much less the Best Buy mouth breathers.
AMD is doing something brilliant here, with the release of Catalyst 8.12 in early December, AMD is going to roll the CAL into the mainstream consumer driver. When you get the graphics drivers, you get GPGPU functionality for free, no download, no install... it just works.
What good is that without software, though? Not much, so that is being addressed as well. AMD defines GPGPU as just about everything not traditionally seen as a GPU task. This means anything other than 3D and video playback is GPGPU. Video encoding, game physics, and science are all classic examples, but you can't help thinking that the killer app is still in someone's basement, being built line by line evenings and weekends.
AMD has all of this, but they aren't being abjectly stupid about it like Nvidia. Over at the other green team, even the cafeteria food has a CUDA tag on it. Trust me, chocolate cake doesn't need to be made with CUDA, and there are some things that even bacon can't make better. AMD is putting things out, Nvidia is banging a drum, take your pick.
A good example of this is the video encoder. Nvidia has been hyping up Badaboom, a custom built app for video encoding. It works great as long as you need to go from one very specific format to another specific format, and quality is not a real concern. ATI's AVIVO encoder is rolled into the 8.12 driver pack, and it is free. Badaboom is $30. And it is slower.
The next gen of tools, SDK 1.3 will be out sometime in Q4, and given that the quarter is half over, this means pretty soon. It will work with all 4xxx series cards, and possibly older versions, but that is not a blanket statement. Stream will be in everything they make going forward, from integrated chipsets to 4870X2s.
Photoshop/CS4 acceleration is there. Silverlight acceleration is there, DRM and all. Office acceleration is there. Cyberlink and Arcsoft acceleration is there as well, and Arcsoft is said to have ported things in about four weeks. In fact, every mainstream app that Nvidia claims support for seems to be supported by Stream as well. Imagine that.
The main catalyst for this is OpenCL, the new GPGPU API that is being worked on by the Khronos group. Think of it as doing for GPGPU what OpenGL did for graphics. It also supports DXnext and OpenGL where applicable, so it is totally standards based. On top of that, you can write whatever you want, the Brook+ language is the one that AMD is pushing, along with several middleware layers, but you can use any OpenCL aware tool.
In the end, the difference is that Nvidia sees GPGPU as a way to drive margins and force people into higher priced cards. ATI is making it ubiquitous and free. Guess which one tends to win in markets, expensive and restricted or cheap and everywhere? If you are still confused, go look up this little company from Redmond called Microsoft...
Should you want the full OpenGL certified drivers for things like AutoCAD and Catia, you can get that too. The new FireStream 9270 is basically a HD4870 with 2GB of DDR5 and a very special driver set. It costs an eye-watering $1,499, but that is about $1,000 less than the competing Nvidia Quadro while packing twice the DP flops. If you need the piece of paper that says it will work with your software, this is your card. For everyone else, it is free. And included.
In the end, AMD/ATI is not being quiet any more. It has the fastest cards on the market, it has a much cheaper professional card, and now it has the APIs. In a few weeks, it will have the CAL everywhere, and what the competition charges an arm and a leg for, it gives out for free.
If Nvidia is counting on its professional line to buoy its bottom line, the next year is going to be ugly. µ
Sign up for INQbot – a weekly roundup of the best from the INQ