Introducing Oracle Acceleron’s multiplanar network. AI workloads demand the largest scale and highest possible network reliability. OCI is addressing these very demands of AI workloads using an innovative network design that relies on multiple parallel network planes.
 


1. Multi-Planar Network  

Multiplanar network uses the concept of disjoint network fabrics (planes) to increase network reliability. Each fabric plane has an independent control plane and independent data plane with no fate shared between planes. This means a network fault in one plane cannot impact other planes – thus reducing the blast radius of network faults. 

Each fabric plane itself is a Clos network. The total bisection bandwidth available to a workload is the sum total of all the planes serving the workload. As a result, network capacity scales with the number of fabric planes – i.e. parallel planes provide yet another scaling multiplier for the AI workload. 

2. High Radix Switches & Shuffle Cables  

Since the scale of the network is determined by the radix of the switch silicon, we use port breakouts to gain an additional scale multiplier for the fabric. A given port (e.g. each 800G port on a 64-port switch) could be split into multiple smaller ports (e.g. 4x200G or 8x100G) to get even larger radix and as a result even larger scale of the network fabric. 

3. Linear Pluggable Optics (LPOs) for Efficiency   

AI workloads are power hungry and sometimes power is the limiting factor in scaling the workload. Network switches and optics consume a notable portion of this power. We use linear optics, such as LPO, to reduce the power consumption of optics by 30%+. Power saved from optics and switches is then allocated to the customer workload – allowing us to serve even larger workloads. 
 
Together, these multi-planar innovations enable OCI to deliver a reliable, scalable, and high-performance network backbone for AI clusters of any size, all while improving efficiency and sustainability. Customers running data-intensive and latency-sensitive workloads can trust OCI’s infrastructure to deliver the scale and reliability required for the next generation of artificial intelligence. 

For more information, see the following resources: