This topic contains 1 reply, has 1 voice, and was last updated by 
Josh Stern August 24, 2023 at 12:41 am.
-
Posts
-
Josh Stern
ModeratorA popular win might include adapters that were not linked to the GPU issue/cost/heat and cheaper implentation of performance like
Switched Fabric – rather than the expense of trying to implement giant, expensive butterfly switches, one can imagine cheap switches with 5,10,20 ports that are dynamically configurable in terms of which link is on. These can then be composed in trees if massive parallelism is required and implemented for the special compute cluster. The software interface would also have a way of dynamically assigning destinations to port/array slots.
You must be logged in to reply to this topic.