Thursday, September 16, 2010

Tunnel Creek

Two days ago Intel announced the E600 system-on-chip (SoC), also known as Tunnel Creek. Intel positions it for in-car infotainment systems, Internet phones, and smart grid devices, but there are good reasons color scientists working on commercial printers should be interested in it.

A system based on an E6xx SoC

The figure above shows a possible system built around an E600 SoC, which is the dark blue part in the top center. The chip includes an Atom core running at 1.6 GHz (E680 part), a GPU running at 400 MHz, and a PCIe interface, drawing only 3.9W of power. The core supports hyper-threading, while the GPU supports OpenGL 2.1; the limitation is that memory is supported only up to 2 GB.

Hyper-threading means that two pages can be ripped concurrently for almost twice the speed, since ripping spends a lot of time on waiting for memory. Not shown in Intel's diagram is that the GPU interfaces to the CPU through the L3 cache. This should give a good speed bump when the rendered page is written back to the CPU for compression.

Two possible architectures spring to mind. The first is to use the system configuration in Intel's diagram above and use their Platform Controller Hub to interface gigabit Ethernet and a disk. The second is a bit more pushed.

Since the E600 has a PCIe interface, a custom hub could be designed, supporting 8 (this is a magic number) E600s, a gigabit Ethernet MAC, a RAID, and the digital press. The RIP would run in parallel using mapReduce, ripping 16 pages concurrently. This would deliver quite a power package costing only about $1000 in volume and drawing less than 100W of power!

And you could put all this on a blade, fill a rack, an use again mapReduce. Will this be a paradigm shift leveraging GPU-based ripping?