What do you benchmark?
NCCL all-reduce and all-to-all across the full cluster against theoretical bandwidth, storage throughput for checkpoints, and failure injection — results delivered with the repository.
Markets · AI labs & GPU clouds
Fabrics measured in NCCL bandwidth, not slide decks.
Model labs, inference providers and GPU clouds need lossless, non-blocking back-end fabrics and a team that benchmarks before handover. Rail-optimized 400/800G RoCEv2 designs, tuned on the twin, delivered as code, with the numbers attached.
What your team will ask us
01
02
InfiniBand or Ethernet?
We build Ethernet. 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.
03
Can you fix an under-performing cluster we already have?
Yes. Most problems are PFC/ECN/DCQCN settings, buffer profiles or rail mistakes; we measure, tune on the twin and re-benchmark.
04
Timeline?
Sizing one to two weeks, design and twin two to four, build dependent on hardware arrival, benchmark and handover one week.
Regulations and how we address them
| Rail-optimized topology | Each GPU rail on its own leaf; non-blocking spine; 400G uplinks today, 800G ready. |
|---|---|
| Lossless Ethernet | PFC, ECN, DCQCN and buffer profiles tuned per platform and validated on the twin. |
| Storage and management fabrics | Separate GPUDirect storage, in-band and out-of-band networks, all as code. |
| Platforms | Arista 7060X/7800R, NVIDIA Spectrum-X, Juniper QFX, Cisco Nexus 9000. |
What you receive
- NCCL benchmark report against theoretical
- Fabric as code with QoS profiles
- Cabling and optics plan certified to 400/800G
- ECN/PFC telemetry in the live portal
Relevant services
Reference design · 10 pages
GPU fabric reference design
Rail-optimized 400G RoCEv2 fabric for 64 nodes / 512 GPUs: topology, oversubscription, QoS and congestion-control parameters, optics and cabling plan, benchmark methodology.