A Novel CMOS ASIC with SPAD Detectors
"CMOS SPADs have many practical advantages over the traditional approaches from other lidar manufacturers. Most important is that they can be directly integrated in a CMOS wafer, which makes it possible to incorporate massive amounts of signal processing on the silicon die right next to the detectors.
As lidar resolutions and data rates continue to increase, on-chip signal processing is essential - the current OS1-64 detector is capable of counting and storing over one trillion photons per second into on-chip memory. This is a titanic amount of data, and we’ve included over 100 GMACs per second (1 GMAC = 1 billion multiply accumulate operations) of signal processing logic in over 10 million transistors to ultimately produce the millions of 3D points per second that our customers use to drive cars, map environments, and identify obstacles.
Processing requirements are only going to go up, and Ouster is leading the way with our custom silicon.
SPADs are also moving along a Moore’s Law-esque performance curve. While today’s SPADs might be 2-5% efficient (efficiency here is measured as the percentage of photons that hit the detector that trigger a binary pulse. Most photons travel through the silicon without causing a pulse to trigger which is a bummer), new SPADs are 20-30% efficient (at 850nm) and efficiencies up to 80% may be possible as the technology matures. Increases in SPAD detector efficiency directly increase a sensor’s range and resolution - a 10x efficiency increase makes an OS1 with 640 lines of resolution not just possible but extremely likely. Unlike the legacy detector technologies used in lidar, our SPAD technology already achieves market leading performance but there is more than an order of magnitude improvement still left to go."
Operating in 850nm band requires more attention to sunlight radiation suppression. Ouster approaches this issue it by concentrating its VCSEL emitter light in sparse points and blocking sunlight from outside of these points, remotely similar to Apple iPad LiDAR:
High sensitive SPADs are mentioned more and more often, but no one talks about their dead time which is equivalently (or more?) important for ranging. A high sensitive and saturated SPAD is useless and no "macro-pixel" can save that.
ReplyDeletewhat do you mean by "no "macro-pixel" can save it?
DeleteMacropixels solve long dead time/ low throughput of the readout. Low SPAD dead time is essential in more and more sensitive SPADs. If a SPAD saturates due to high PDE and long dead time, it is useless. Aggregating 2 or more saturated SPADs gives us saturated macro-pixels.
ReplyDelete