News Nvidia-backed startup invents Ethernet memory pool to help power AI — claims it can add up to 18TB of DDR5 capacity for large-scale inference workl...

[Admin]

Enfabrica is piloting its Emfasys Ethernet-connected DDR5 memory pool that adds up to 18TB of CXL-attached memory to any AI server to ease memory bottlenecks in large-scale inference workloads.

Nvidia-backed startup invents Ethernet memory pool to help power AI — claims it can add up to 18TB of DDR5 capacity for large-scale inference workl... : Read more

 

[rluker5]

Latency in the microseconds range though. Shame they can't just use Optane for this.

 

[bit_user]

Can't they just use CXL without layering it atop Ethernet? Ethernet certainly scales better, but it burns a lot of power and adds nontrivial amounts of latecy (as @rluker5 mentioned).