AI BriefingRebellions Engineering: Switchless Dragonfly and 6D Hypercube Topologies Offer Cost-Effective 64-Device AI Racks
·2026.10.21 22:00
Key point
Rebellions AI offers switchless Dragonfly and 6D hypercube topologies for Rebel 100s clusters, reducing capital and power costs while maintaining competitive performance for large-scale GenAI inference.
Details
- Rebellions AI provides flexible network topologies (Dragonfly, 3D/6D hypercube) for GenAI clusters, contrasting with Nvidia's fixed All-to-All NVSwitch architecture.
- Benchmarks show that adding Ethernet switches to interlink 64 Rebel 100s accelerators increases inference throughput by up to 52.4% and interactivity by 47.6% compared to the switchless Dragonfly baseline.
- The switchless Dragonfly and 6D hypercube topologies allow for a 64-device rack without external switches, reducing capital, power, and space expenses to lower cost per token.
- Tests used a 3-trillion parameter mixture-of-experts model with FP4-quantized weights, requiring a minimum of 10.9 Rebel 100s accelerators and 16 XPUs for data parallelism.
- Rebel 100s integrated routers currently have a latency of around 1 microsecond, with Rebellions working to reduce this to the ~500 nanoseconds estimated for Nvidia NVSwitch, while providing 1.6 TB/sec aggregate bandwidth across eight Ethernet ports.