I’m tracking the $MRVL and $PENG thread: the sample delivery target for optical interconnect servers in the second half of 2027 points to a much bigger shift—memory can now scale independently of compute.
Today, if a workload needs more memory, you basically have to add another entire server, driving up processor, networking, and power requirements all at once.
Shared memory offers a different way to solve the problem.
Today, if a workload needs more memory, you basically have to add another entire server, driving up processor, networking, and power requirements all at once.
Shared memory offers a different way to solve the problem.