@andymilkis happy to take part in that stream if a non-artist perspective is of interest?
Geeky Flint(RT) stuff: the Indigo2 Impact (the magenta one) still only had 2 x Fast/Narrow SCSI busses (one of which was internal, so we had that ugly cable/bracket to bring it out to the back of the machine), which limited you to around 15MB/sec of disk I/O (using a small 4 drive array). Y’CbCr 4:2:2 8 bit D1 video is 20MB/sec, RGB 4:4:4 is 30 MB/sec, so not enough disk I/O for real time capture/playback. Also the MIPS R4400 systems were limited to 384MB of RAM (the later R10K model could take more, but for some reason I don’t remember I don’t think we ended up supporting the R10K models).
We implemented “burst mode” video capture: we would capture to memory in Y’CbCR 4:2:2 while trying to drain to disk as fast as possible, but at some point the memory would fill up and we’d have to stop and wait for it to drain to disk, don’t remember the how long a burst you would get but around 15-20 seconds maybe? And once your clip was fully captured, we would go back and post-process to convert to RGB 4:4:4 in non-realtime. Similarly for output we would first pre-process from RGB 4:4:4 to Y’CbCr 4:2:2, pre load as much in memory as possible, and then output in chunks, emptying the memory buffer faster than it could be filled by limited disk bandwidth.
Full res playback could only happen from memory, or in proxy res from disk.
FlintRT was an attempt to offload real-time capture, output and playback to a dedicated DDR called Pebbles. Unfortunately a decision was made to use a type of IBM disks which used a novel serial interface called SSA (Serial Storage Architecture, long before SATA or SAS was a thing), and those drives had terrible reliability and performance problems, which hobbled Pebbles.
By the time these hardware problems got resolved, the Octane was on the horizon. Most importantly, Octane had more than enough I/O to do D1 RGB 4:4:4 I/O without standing on your head: it had an external Ultra/Wide/Single Ended SCSI port which could do 35MB/sec by itself, and could take the 4 port XIO Ultra/Wide/Diff SCSI card or the 2 port XIO 1gbit/sec copper FC card, which eventually proved even fast enough for HD I/O on the Octane 2.
It was a blessing and a curse to be so closely tied to the SGI platforms: it let the software do things that other systems typically could not at the time, but it also drove some product decisions that didn’t always make clients happy. From the perspective of a developer, we always tried to do the best we could with the platform.
That’s great stuff, man. Thank you for sharing!
