ARM Community: GPU Computing, the OpenCL debate and performance measurement - ARM Community

Recently, a lot of airline miles have been flown between my team and ARM partners to engage in conversations about the latest and future uses of GPUs. Mostly, I have been discussing our graphics roadmap which I still can’t tell you about publicly, but I also got a chance to see some of the latest cool demos on our current GPUs. I also got a chance to hear from our partners – about the slew of consumer products that are coming out soon based on our technology, and about the things they are doing with our GPUs. These partners, a fantastic collection of the most important companies in our industry, really are doing it all: printers, PNDs, digital still cameras, digital TV, set-top-box, automotive, mobile computing, and mobile handsets. You name it, they are doing it; if it has a screen, they are using GPUs to make them look more attractive and while these are great times, there are even better to come when these hit the shops.

The OpenCL debate in my team
Recently, the really interesting discussions from a technological standpoint have been about the use cases for GPU computing. Tom Olson has had some fun recently in his OpenCL blog: where he tries to pick a fight with me. I shall be smug and superior and refuse to rise to such a challenge! Also, without being a weasel, we’re both right! The more people we talk to, the more use cases we find...

GPUs accelerate driver assistance and much, much more
The automotive boys want to fit a GPU for exciting, but intuitive displays. They also want to use GPU computing for driver assistance. Think about a camera and image analysis that can read road signs to superimpose the speed limit on the speedometer display, or warn if it detects children running out into the road. White line tracking and lane detection is being done now, but they also look like a good candidate for acceleration through GPU computing. In cost-sensitive devices (and that's just about everyone these days) partners are looking at the BOM to see if a highly-capable GPU supporting OpenCL can help them remove other computational devices like some of the DSPs in complicated SoC designs. The applications being run on these DSPs today are many and varied: some look good for a GPU and some less so, but time will tell. The no-brainers are the ones being done where an algorithm is applied over large datasets - the same routine being applied to a vast array of data, for example. I've come across several of these; some in the obvious fields of image computation (colour conversion for example), but others not so obvious in more numerical applications, some of which don't typically even have screens!

GPU vs. CPU or DSP: Beware of what you are measuring
Note to system architects: one common theme from my discussions was that graphics cards for desktop machines weren't very useful for experimentation in this area. More than one OEM complained about the system architecture of these devices not being suitable - where the "GPU memory" and the "CPU memory" address spaces were separate, and where they were physically separate memory devices as well. What everyone wants to do is to try out their algorithms on such platforms to find out how fast they will go on a GPU rather than CPU or DSP in an embedded GPU environment. What they are currently tending to find is that they end up measuring the performance of the bridges connecting the different memories - not what they wanted at all. Efficient (in bandwidth and power terms) connections between CPU and GPU will be critical in these new devices. GPU designers that can utilise the memory architecture (buses, paths to CPU, paths to memory controllers) efficiently (e.g. long bursts, efficient cache-to-cache coherency) will be able to differentiate their products and customers using that IP will gain advantages over their competitors.

Are you an OEM just itching to get your hands on chips supporting OpenCL and DirectCompute? Tell us about your applications.

Jem is an ARM Fellow and likes to think of himself as "The Godfather" to technical talent in ARM. After spending some time in his youth writing software for satellites and traffic-lights among other fascinating things, Jem spotted the technical inflection point of the mobile industry: graphics, video and other visual processing. As VP of technology in the Media Processing Division of ARM, Jem is busy with a lot of projects involving the future of cool ARM technology, which will revolutionise how people experience and interact with digital devices.

Shortlink to this post: http://bit.ly/a5G4S2