Deliver low-latency, low-power, high-bandwidth-density near-packaged optical engines required to address the extended reach connectivity needs for multi-rack-scale GPU clusters.
Challenges
- High-radix switch and compute architectures need high-bandwidth-density optical connectivity across multiple racks to link GPU clusters of increasing size
- To maximize GPU utilization and improve token throughput, AI workloads demand low-latency optical engines with improved signal integrity
- As optics are increasingly pulled into the scale-up domain, power budget becomes a bottleneck that designers must accommodate with low-power component choices where possible
- Hyperscalers require a viable manufacturing ecosystem and reliable serviceability to deploy high-radix optical connectivity at scale today
- Maintaining uptime for large GPU clusters is difficult because a single failed link can disrupt distributed workloads and reduce utilization across the entire system
Solution: Near-Packaged Optics (NPO) With Aries PCIe® Smart Optical Drivers and TIAs
- High-bandwidth-density NPO with smart optical drivers and TIAs enable links extending beyond 50 meters across multi-rack systems
- High-density NPO with linear optical engines placed next to the switch and compute chip enable improved signal integrity and provide low-latency fabric connectivity
- NPO with low-power smart optical drivers and TIAs provide the extended reach across multiple racks while minimizing the overall system power consumption
- Electrically pluggable NPO leverages the same robust, proven manufacturing and test ecosystem as front panel pluggables, enabling hyperscalers to deploy optical connectivity at scale
- COSMOS Link Intelligence improves system resilience by providing telemetry, link monitoring, and dynamic tuning of purpose-built NPO with Smart Optical Drivers and TIAs, helping detect and resolve connectivity issues before they cascade across the GPU cluster
