MonolithicIC3D Presents 3 Sensors, ROIC, ISP and Memory Integration
At Semicon West on July 12-14, 2011 MonolythIC3D is going to present a technological possibility of integration of 3 sensor layers, ROIC, ISP and memory in a single package:
Standard DARPA-like concept chart. Really, who needs this? Maybe for an eyeball at the front of a guided bullet? SIMD architectures with a detector in each PE has been conceptually around for at least 30 years* (because it was in my Ph.D. thesis, and it wasn't new then, and also in my student Sayed Eid's thesis, and keeps popping up from CS faculty across the globe) and yet it has never found a good application for monolithic implementation. I don't think 3D vertical stacking is going to help much for adoption.
SIMD is good for a lot of stuff, but not for higher level vision at all. And with small pixels, how much of a processor can you afford? And if they are few in number and big pitch, who needs a parallel processor? Eid's chip, by the way, had 4 detectors per PE.
But who knows, maybe the time has come for this concept?
*for example, see "Architectures for Focal-Plane Image Processing" http://ericfossum.com/Publications/Papers/Architectures%20for%20Focal-Plane%20Image%20Processing.pdf
Indeed, I worked on this during my PhD thesis too. People focused only on data parallelism in individual data crunching, but forgot the key issue in a SIMD: data communications. When the processing goes to higher and higher level, the communications become more and more global and regular grid like that in most of SIMD becomes very poor in performance. The best example is the Connection Machine 2, the global communications are so slow... By the way, does Thinking Machine still exist? Anyone knows ?
For small pixel array, the standard digital processing is a good solution and the shared memory gives a unity distance communication for any connection pattern.
But I do think that in pixel intelligent can play an important role in some fast reflex based actions. We have worked on in-pixel histogram based adaptive sensor, it works pretty well.
The question now is that the development of CMOS sensor is so expensive that we have to find some sounding and resonable volume applications for such idea, that is a challenge ! Maybe the homing sensor in a bullet is good one :)
"Histogram-equalization-based adaptive image sensor for real-time vision" IEEE SSC Journal, Jul. 1997, Issue 7, pp. 1027-1036
Standard DARPA-like concept chart. Really, who needs this? Maybe for an eyeball at the front of a guided bullet? SIMD architectures with a detector in each PE has been conceptually around for at least 30 years* (because it was in my Ph.D. thesis, and it wasn't new then, and also in my student Sayed Eid's thesis, and keeps popping up from CS faculty across the globe) and yet it has never found a good application for monolithic implementation. I don't think 3D vertical stacking is going to help much for adoption.
ReplyDeleteSIMD is good for a lot of stuff, but not for higher level vision at all. And with small pixels, how much of a processor can you afford? And if they are few in number and big pitch, who needs a parallel processor? Eid's chip, by the way, had 4 detectors per PE.
But who knows, maybe the time has come for this concept?
*for example, see "Architectures for Focal-Plane Image Processing"
http://ericfossum.com/Publications/Papers/Architectures%20for%20Focal-Plane%20Image%20Processing.pdf
Indeed, I worked on this during my PhD thesis too. People focused only on data parallelism in individual data crunching, but forgot the key issue in a SIMD: data communications. When the processing goes to higher and higher level, the communications become more and more global and regular grid like that in most of SIMD becomes very poor in performance. The best example is the Connection Machine 2, the global communications are so slow... By the way, does Thinking Machine still exist? Anyone knows ?
ReplyDeleteFor small pixel array, the standard digital processing is a good solution and the shared memory gives a unity distance communication for any connection pattern.
But I do think that in pixel intelligent can play an important role in some fast reflex based actions. We have worked on in-pixel histogram based adaptive sensor, it works pretty well.
The question now is that the development of CMOS sensor is so expensive that we have to find some sounding and resonable volume applications for such idea, that is a challenge ! Maybe the homing sensor in a bullet is good one :)
"Histogram-equalization-based adaptive image sensor for real-time vision"
IEEE SSC Journal, Jul. 1997, Issue 7, pp. 1027-1036
-yang ni