NUMAlink 2 was designed for use in SGI’s Origin® 200 and
Origin® 2000 family of compute servers. At the physical layer,
NUMAlink 2 operates at a switching rate of 400Mb/s across
20 data bits/direction. The NUMAlink 2 bandwidth is 0.8GB/s/
direction. A source synchronous clock is utilized to capture
data in parallel at the receive end in favor of higher overhead
clock encoding/extraction schemes. The SERDES (serializer/
deserializer) functionality was optimized for minimal latency.
One example is in the serializer section where computed link
layer sideband bits can be latched late in the serialization cycle
rather than having to wait until the beginning of the next cycle.
The physical layer supports up to 24" of FR4 printed circuit
board (PCB) interconnect, 3 meters of twin-ax based cables
and organic ASIC packages. The NUMAlink 2 physical layer
latency including 24" of PCB and 3 meters of cable totals 55 ns.
The link layer features the robust 16 bit CRC-CCITT (a standard
algorithm) to protect against data errors and all detected errors
are corrected through a Go back N Automatic Repeat Request
protocol. Efficient 2 level router look-up tables are used to route
local and global routes. A 6 port NUMAlink 2 router was
designed to achieve processor scalability up to 512 sockets in
a globally addressable, single system image.
No comments:
Post a Comment