Workload-to-topology sizing
GPU count, collective patterns and model sizes translated into rail-optimized leaf-spine with the right oversubscription — usually none.
Signature program
Built for GPUs. Lossless. Non-blocking.
Rail-optimized 400G and 800G Ethernet fabrics with RoCEv2, designed around your training and inference workloads — PFC, ECN and QoS tuned on the twin, NCCL benchmarks run at handover, and the fabric delivered as code.
What's included
01
02
Lossless Ethernet design
RoCEv2 with PFC, ECN and DCQCN parameters, buffer profiles and QoS classes tuned per platform and proven on the twin.
03
Platform selection
Arista 7060X/7800R, NVIDIA Spectrum-X, Juniper QFX or Cisco Nexus 9000 — chosen for the workload, with 800G and 1.6T-ready optics.
04
Front-end, storage and management networks
Separate fabrics for storage (GPUDirect), in-band management and out-of-band, each as code.
05
Validation
NCCL all-reduce and all-to-all benchmarks across the full cluster, failure injection, and a performance baseline you keep.
06
Operations
Telemetry for ECN marks, PFC pauses and buffer occupancy in the live portal; optional AIOps operations.
How an engagement runs
01
Sizing
Workload interviews and cluster plan. One to two weeks.
02
Design
Topology, QoS, optics, cabling, as code; twin validation.
03
Build
Rack-and-stack, cabling certification, bring-up.
04
Benchmark and hand over
NCCL results against targets, documentation, repository.
Questions we get asked
InfiniBand or Ethernet?
We build Ethernet fabrics. For most clusters under a few thousand GPUs, a well-tuned RoCEv2 fabric reaches the same job performance with simpler operations and multi-vendor choice.
Can you tune a fabric someone else built?
Yes. Most under-performing training clusters trace back to QoS and congestion-control settings we can measure and fix without new hardware.
Request a quote
Tell us about the project; a senior engineer responds the same business day.