Prefill-Decode Disaggregation

Separating the prefill (prompt processing) and decode (token generation) phases of LLM inference onto different hardware, because they have fundamentally different compute profiles.

Why It Matters

  • Prefill is compute-intensive (matrix multiplications over the full prompt)
  • Decode is memory-bandwidth-intensive (one token at a time, need to read all KV cache)
  • Running both on same hardware means one phase is always underutilized

Evolution

  1. Single-cluster PD: Standard approach (Mooncake, vLLM, SGLang). Prefill and decode on separate nodes within same RDMA fabric.
  2. Cross-datacenter PD (PrfaaS): Selective offloading of long prefills to compute-dense clusters, KVCache over commodity Ethernet. Enabled by hybrid attention architectures reducing KV by ~10x.

Hardware Divergence

  • Prefill-optimized: NVIDIA Rubin CPX (high compute throughput)
  • Decode-optimized: Groq LPU (extreme memory bandwidth)
  • These can’t easily share the same RDMA fabric → need cross-datacenter solutions

Key Constraint

The “bandwidth wall”: for dense-attention models, a single 32K request generates ~60 Gbps of KVCache traffic. Hybrid attention (Kimi Delta Attention, sliding window) reduces this by 10x, making cross-datacenter transfer feasible.

Connections