# Aviz Networks and Marvell discuss production-ready Sonic networks

URL: https://decisioninsights.ai/aviz-networks-and-marvell-discuss-production-ready-sonic-networks/

Aviz and Marvell describe a partnership pairing Marvell silicon with Aviz services to deliver production-ready Sonic networks. The speakers link AI-driven operations to reduced GPU idle time, a point relevant to enterprise IT leaders.Partnership BackgroundMarvell said it partnered with Aviz after acquiring Innovium, which had offered a certified Sonic program to its customers. Marvell said it preferred to focus on silicon rather than OEM business, and Aviz began deployments with it from the start and took over existing customer deployments, including eBay.Division of ResponsibilitiesMarvell described its role as providing silicon and system designs, with ODM partners building the systems. The speakers said Aviz work begins at day minus one, when customer requirements are modeled and tested before day zero deployment. Production readiness was described as covering hardware validation, feature testing, automation, and ongoing support such as bug fixes and added telemetry.Marvell Silicon and AI NetworkingMarvell said its Innovium silicon was designed for data center environments, which it said provides lower latency, lower power, and flexible pipelines. The company said its silicon has been deployed in production by hyperscalers and enterprises. Marvell also said it supplies optical connectivity technologies and works with hyperscalers on custom XPU designs.Open Networking and SonicThe speakers described open networking as a model in which the same software layers run across different hardware, and referred to Sonic as open networking 2.0. Microsoft open sourced Sonic around 2017. The speakers said hyperscalers have the staff to operationalize Sonic internally, while enterprises need vendor support to make it production ready.AI and GPU Cluster NetworkingThe speakers said AI workloads shifted networking from a single front-end network to a larger back-end network connecting GPUs, with scale-up and scale-out designs growing from 8 or 16 GPUs per cluster to 72 and beyond. Latency, power, and avoiding packet drops were cited as increasingly important. Idle GPU time was described as costly, since GPUs waiting on data consume power and represent lost value.AI-Driven Network OperationsThe speakers said AI operations layered on existing infrastructure could speed troubleshooting, simplify management, and support predictive analysis, such as identifying growing queue sizes on switches before failures occur. They said this could reduce downtime and the resources spent on troubleshooting and monitoring. The approach depends on ingesting network data into models.Partnership OutlookThe speakers said the partnership aims to extend OEM-style support to more enterprises at lower cost, combining Aviz capabilities with Marvell silicon. They cited total cost of ownership reductions, including operational savings that accumulate over years.This Blog Summary is a fact-based summary of the vendor podcast transcript. Enterprise decision-makers evaluating open networking, Sonic production readiness, and AI cluster infrastructure may find the stated division of responsibilities and cost points relevant to planning.
