...

• Need to buy a big node (like hsd node) to hold 6 kcu cards
• dionisio says this mtp24 is only available in multimode:  https://www.mouser.com/ProductDetail/Amphenol-FSI/10124588-310?qs=81r%252BiQLm7BSk9m%252B2puhj7Q%3D%3D.
• see also https://www.amphenol-cs.com/product-series/leap-on-board-transceiver.html
• see also https://www.amphenol-cs.com/optics/transceivers.html
• amphenol tells us to contact our local sales rep:  Gaurav.Singh@amphenol-fci.com
• feedthroughs may also be a problem.  Kaz is getting part number RETA-RIB-CF40-24-GI50-025-S70-MTPF (24 fibers) and RETA-RIB-CF16-12-GI50-025-S48-MTPF (12 fibers) from sedi-ati.com.  These two parts are similar to RETA-STANF21.208 (different length) and RETA-LBNL19.141 (different flange).  Looks like feedthroughs are available in single-mode!
• transceiver outputs MPO24 (12 fiber pairs).  the vacuum feedthroughs will also be MPO24
  • timing comes from XPM.  configuration/monitoring/data fibers connect to DRP.  we will use LC plus cassettes to get to MPO24 (and MPO8 going to the multimode to single mode converters).
• need to have/order:
  • a node to hold all the multimode-to-single-mode converter KCUs
  • LC to MPO24 cassettes
  • MPO8 breakout cables

Amphenol says these are "variants" of the transceiver that they support:

SZ Compression

Chuck writes about roibinsz:

...

This 71ms number + 20ms for calibration feels tight on 10 nodes.  Implies 10Hz per core, or 500 cores (8 nodes)

This paper compares CPU/GPU SZ compression:  https://arxiv.org/abs/2007.09625.  It states 370x improvement on GPU vs 1 CPU core.  If we have 60 usable cores on a DRP node, that makes a GPU-SZ node the equivalent of 6x faster than a CPU-SZ node.

Schedule

(on Dec. 19, 2022) Sometime in February/March a prototype with 3 tiles will be released (could generate fake data for missing panels for 1 quad).  Full detector: summer?

...