Ethernet or InfiniBand?

We design lossless Ethernet (RoCEv2) on open hardware — the approach Edgecore's public AI deployments run; the trade-offs are stated openly in the design.

Will it work with our NVIDIA/AMD servers?

The fabric is server-vendor-neutral; NIC models and firmware are part of the design inputs.

We do, and the acceptance test publishes measured results against agreed targets before handover.

Yes — the reference design is pod-based; a 2-rack pod carries the same architecture as a 20-rack build.

Stated per SKU at BOM; the design offers alternates where supply differs.

Support after handover?

Runbook-based support tiers up to managed fabric operation from Bangkok.

HOME/SOLUTIONS/AI Fabric

AI FABRIC

The GPU servers arrived; the switching didn't keep up. Training jobs stall on communication, storage can't feed the cluster, and the next expansion doubles the problem. You need a lossless Ethernet fabric engineered for the cluster you actually run.

WHAT YOU WILL HAVE AT THE END

Non-blocking 100/400/800G fabric sized to your GPU count
Lossless RoCEv2 transport — PFC/ECN engineered end to end
Storage fabric that feeds the cluster at training speed
A design that scales to the next GPU tranche without re-architecture
Telemetry per port and per queue from day one
Documented fabric you can reason about at 2 a.m.

REFERENCE DESIGN

REFERENCE DESIGN (PDF) ↓

Leaf / spineAS7726-32X (100G) · DCS-series 400G platforms400/800G SKUs confirmed per cluster size at design time

NOSSONiC (Enterprise or Community)

Optics / DACET7402-SR4/CWDM4 (100G) · ET7302 (25G server)800G optics specified at design from Edgecore's transceiver line

WHAT WE DELIVER

01Cluster workload profile and fabric sizing
02Engineered design + BOM
03Hardware supply
04Fabric bring-up incl. RoCE tuning
05Acceptance: measured all-to-all performance against the design targets
06Runbook + documentation handover
07Support tiers

SIZING & TIMELINE

BANDSCALETIMELINE

S≤64 GPUs, 1–4 racks3–5 weeks

M≤512 GPUs6–10 weeks

L512+ GPUs, multi-pod10–16 weeks, phased

Outright purchase
เช่าใช้ 36-month service agreement
Design-only engagement if procurement runs elsewhere

PROVEN IN THE FIELD

Edgecore SONiC fabrics carry production AI training workloads today. Photon State designs and delivers this solution in Thailand on the same platform.

SAKURA Internet (SAKURAONE)

JAPAN · 2025

SAKURA Internet built its 800-GPU SAKURAONE AI cluster on Edgecore AIS800-64O 800GbE switches running SONiC, arranged as a rail-optimized leaf-spine fabric with RoCEv2 in place of InfiniBand. The system reached #49 on the June 2025 Top500 as the only top-100 machine on a fully open Ethernet networking stack, published by SAKURA's own researchers at MLSys 2026.

#49TOP500June 2025 list 33.95PFLOP/SHPL Rmax

HyperVerge

INDIA · 2026

AI verification company HyperVerge deployed Edgecore high-bandwidth data center switches running SONiC to interconnect its GPU training cluster and a clustered NVMe parallel file system for face-recognition, OCR, and custom LLM workloads. The company reports its 100G RoCEv2 lossless fabric delivers the pooled storage at close to local-NVMe latency, and that open standards let its team self-configure and monitor the network quickly.

QUESTIONS BUYERS ASK

Ethernet or InfiniBand?: We design lossless Ethernet (RoCEv2) on open hardware — the approach Edgecore's public AI deployments run; the trade-offs are stated openly in the design.
Will it work with our NVIDIA/AMD servers?: The fabric is server-vendor-neutral; NIC models and firmware are part of the design inputs.
Who tunes PFC/ECN?: We do, and the acceptance test publishes measured results against agreed targets before handover.
Can we start small?: Yes — the reference design is pod-based; a 2-rack pod carries the same architecture as a 20-rack build.
Lead times on 400G?: Stated per SKU at BOM; the design offers alternates where supply differs.
Support after handover?: Runbook-based support tiers up to managed fabric operation from Bangkok.

Request a free AI fabric consultation →