RoCE Ethernet Switches
RoCEv2-capable data centre switches for AI and storage networks. Lossless fabric with PFC and ECN, spine-leaf architecture, open network operating system options.
RoCE (RDMA over Converged Ethernet) delivers RDMA’s low-latency, CPU-bypassing data movement over standard Ethernet infrastructure. The price of that is a requirement: the network must be configured to run lossless. When priority flow control (PFC) and explicit congestion notification (ECN) are not tuned correctly, RoCE falls well short of expected performance.
The data centre switches in our catalogue serve spine-leaf architectures with 800G, 400G and 100G port options. The open networking approach lets hardware and network operating system be chosen separately, which reduces single-vendor lock-in and fits your automation tooling.
Mevasis does not just supply RoCE fabrics; we apply the lossless configuration and verify it with measurements after commissioning.
How to choose
Switching capacity and port density
Does the switch's total capacity support every port running at full rate simultaneously? At the spine layer, one high-capacity box means fewer hops and a simpler topology than several smaller switches. Higher port density also lowers rack and cabling cost.
Lossless transport capabilities
What determines RoCEv2 performance is congestion management, not raw speed. PFC and ECN support, buffer architecture and dynamic load balancing keep GPUs from waiting on packet loss. These capabilities are listed explicitly in the specification tables on product pages.
Network operating system
On open networking switches, hardware and software are decoupled: the same hardware can run different network operating systems. Choosing one that fits your team's CLI and automation tooling is what determines operating cost over the long run.
Separating storage from compute
Carrying storage traffic on the same plane as compute traffic introduces unpredictable latency into training jobs. A dedicated storage plane costs extra switches but keeps both sides predictable.
Products 2
FS
PicOS N8560-32C — 32 x 100G Data Centre Switch
32 x 100G/40G QSFP28 ports — up to 128 x 25G/10G with breakout3.2 / 6.4 Tb/s switching capacity (unidirectional / bidirectional)Broadcom Trident 3-X7 silicon in a 1U chassis
View details
Price on request
FS
PicOS N8650-32OD — 32 x 800G AI/HPC Switch
32 x 800G OSFP ports, 25.6 Tb/s switching capacity (1U)A 64-port, 51.2 Tb/s variant (2U) in the same familyBroadcom Tomahawk 5 silicon
View details
Price on requestFrequently Asked Questions
Both deliver RDMA. InfiniBand provides lossless transport by design and needs less tuning, which puts it ahead in pure HPC clusters chasing the lowest latency. RoCE leans on the existing Ethernet ecosystem, your team's skills and multi-vendor choice. For storage networks and mixed workloads RoCE is often more practical. Share your workload profile and we will provide a comparative assessment.
PFC priority classes, ECN thresholds and buffer settings are part of our commissioning service. We frequently see fabrics left at defaults failing to deliver the performance RoCE is capable of.
Technically yes, provided the spine switches support lossless transport. Share your current switch models and we will report which part can be reused and which needs replacing.
Let's Build Future Together.
