Taylor: Avoiding Jitter in Composited Frame Display
Posted Dec 2, 2012 10:42 UTC (Sun) by
farnz (guest, #17727)
In reply to:
Taylor: Avoiding Jitter in Composited Frame Display by Jonno
Parent article:
Taylor: Avoiding Jitter in Composited Frame Display
Note that the driver for high frame rates on the link is 3D TV - the killer is the link bandwidth to transfer a high refresh rate image. Many screens are already capable of 300 Hz or 600 Hz refresh, if only you could get image data to the panel that quickly.
Standard TV-style 1080p60 has a pixel clock of 148.5 MHz - 3D TV uses tricks similar to old-style interlace to get "120 Hz" refresh in that, by effectively transmitting two 74.25 Mhz pictures (2x1080i60 for example). If we wanted to send 300 Hz down the wire, we need a pixel clock of 742.5 MHz. Note too that we send 24 bits per pixel, for 17.82 GBit/s data to shift on a 1080p300 display. Add in the fact that "retina" type displays have 4 times the pixels, and a single 3840x2160p300 display needs you to transfer 71.28 GBit/s just for scanout.
Note that 71.28 GBit/s is the lower bound there; that's just the RAM throughput needed to drive a 300 Hz retina display, assuming a static image. For comparison, a desktop IvyBridge CPU from Intel has 25.6 GByte/s peak memory bandwidth for all uses; that's 204.8 GBit/s in the best scenario. We're talking about eating 1/3rd of that for scanout, assuming the memory controller can sustain the peak. In practice, you normally see around half the peak rate as sustained data rate with reasonably optimized memory access pattern - we're now talking about eating 2/3rds of the available memory bandwidth just to scan out the image, leaving us with very little memory bandwidth left to render it.
Bear in mind too that graphics is often memory-bandwidth limited; one of the things you're paying for when you buy a $600 card rather than a $50 card or integrated graphics is a very high bandwidth memory bus. Eating lots of that memory bandwidth for scanout is going to adversely affect performance of the rest of the GPU's operation.
(
Log in to post comments)