Speaker: David Luebke (University of Virginia)
The ultimate display will not show images. To drive the display of the future, we must abandon our traditional concepts of pixels, and of images as grids of coherent pixels, and of imagery as a sequence of images. So what is this ultimate display? One thing is obvious: the display of the future will have incredibly high resolution. A typical monitor today has 100 dpi-far below a satisfactory printer. Several technologies offer the prospect of much higher resolutions; even today you can buy a 300 dpi e-book. Accounting for hyperacuity, one can make the argument that a "perfect" desktop-sized monitor would require about 6000 dpi-call it 11 gigapixels. Even if we don't seek a perfect monitor, we do want large displays. The very walls of our offices should be active display surfaces, addressable to a resolution comparable to or better than current monitors. It's not just spatial resolution, either. We need higher temporal resolution: hardcore gamers already use single buffering to reduce delays. The human factors literature justifies this: even 15 ms of delay can harm task performance. Exotic technologies (holographic, autostereoscopic...) just increase the spatial, temporal, and directional resolution required. Suppose we settle for 1 gigapixel displays that can refresh at 240 Hz-roughly 4000x typical display bandwidths today. Recomputing and refreshing every pixel every time is a Bad Idea, for power and thermal reasons if nothing else. We will present an alternative: discard the frame. Send the display streams of samples (location+color) instead of sequences of images. Build hardware into the display to buffer and reconstruct images from these samples. Exploit temporal coherence: send samples less often where imagery is changing slowly. Exploit spatial coherence: send fewer samples where imagery is low-frequency. Without the rigid sampling patterns of framed renderers,sampling and reconstruction can adapt with very fine granularity to spatio-temporal image change. Sampling uses closed-loop feedback to guide sampling toward edges or motion in the image. A temporally deep buffer stores all the samples created over a short time interval for use in reconstruction. Reconstruction responds both to sampling density and spatio-temporal color gradients. We argue that this will reduce bandwidth requirements by 1-2 orders of magnitude, and show results from our preliminary experiments.