Dell offers scalable, high-density 400GbE/100GbE fabrics using the Dell PowerSwitch Z9664F-ON.
This switch is recommended for robust AI distributed training applications. The Z9664F-ON offers:
- AI/ML Enablement – RoCEv2-based AI Fabric at GPU scale with a single switch.
- 200G – allows rollout of RoCEv2-based Fabric at 128 GPU scale with a single switch.
- 400G – (future) allows rollout of RoCEv2-based AI fabric with 64 GPU scale.
- High Radix/High density: 100 GbE (256x100G) platform for scale-out Layer 3 fabrics in super spine architecture.
- Middle of Row: Directly connect to multiple rack servers with a single switch.
- SAN Switching: Direct connect to up to 256 nodes—fewer layers means better latency and less complexity.
- Usable as a 10/25/50/100/400G switch
- Dell Enterprise SONiC NOS (See Dell Enterprise SONiC Specifications and Features Matrix)
Z9664F-ON switches can be used for virtually any AI workload environment:
AI Workload | Characteristics |
Training | Sustained "elephant or long," bursty, and nonlatency sensitive traffic flows that take place in large clusters with predictable traffic patterns. |
Fine-tuning | Like training but at a smaller scale (as discussed in this paper). |
Inferencing | Latency-sensitive and unpredictable traffic patterns. |
Z9664F-ON switches will also scale as your AI network requirements expand. See the example topology in Leaf-spine below to find out how far.