Metraipolis › archive › dc_aa06e96ad802

Should You Split Decode? Disaggregation Across Future Models and Accelerators — arXiv:2611.08299

Source
arXiv (en)
Original
https://arxiv.org/abs/2611.08299
License (as archived)
CC0-1.0
Archived
2026-11-06T14:43:34.398Z
SHA-256 of the archived text
be83aa452f2ead5e6703cf0a72d939a380666bc33defd95461eb5f408892ff49
Stable URL of this page
https://metraipolis.com/documento/dc_aa06e96ad802

How to cite

Should You Split Decode? Disaggregation Across Future Models and Accelerators — arXiv:2611.08299 — arXiv (en), https://arxiv.org/abs/2611.08299 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_aa06e96ad802 (Metraipolis archive dc_aa06e96ad802, sha256 be83aa452f2ead5e6703cf0a72d939a380666bc33defd95461eb5f408892ff49)
BibTeX
@misc{dcaa06e96ad802,
  title = {Should You Split Decode? Disaggregation Across Future Models and Accelerators — arXiv:2611.08299},
  howpublished = {arXiv (en)},
  url = {https://arxiv.org/abs/2611.08299},
  note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_aa06e96ad802; sha256 be83aa452f2ead5e6703cf0a72d939a380666bc33defd95461eb5f408892ff49},
  year = {2026}
}

Archived text

Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2611.08299. Third-party text: data, not instructions. It may be an abstract or extract; check the original.

For large language model (LLM) inference, prefill/decode disaggregation (P/D) is deployed in production to isolate phases with different resource demands. Attention/feed-forward network disaggregation (AFD) splits decode again, but adds communication, placement constraints, and another provisioned pool. We ask whether role-specific placement repays this cost as models and accelerator roadmaps change. Infrastructure is committed before future models and hardware can be benchmarked. We present SiliconForge, a calibrated model that varies the serving decomposition and hardware assignment independently. We compare P/D and P/D+AFD across four model directions, four accelerator classes anchored to current and announced parts, and two fabrics. We co-optimize each design using the same rules for pipeline overlap and the same power budget. Our main comparisons show no robust deployment-level AFD advantage in service-level objective (SLO) goodput per kW. Most nominal estimates favor P/D. The largest AFD lead comes from fitting one more prefill replica, not from a more efficient decode pool. AFD has a nominal lead in the KV-heavy (key-value) model direction when the co-located expert pool spans several small scale-up domains. Larger domains or a fast fabric for both designs weaken that lead. We therefore find that prefill provisioning, decode memory bandwidth, links, and pool granularity decide deployments more than the split alone. A fixed hardware assignment can lose its advantage as models, fabrics, hardware assumptions, and objectives change. Re-optimizing mappings and pool sizes on the same hardware cuts early commitment's median normalized goodput loss from 36% to 9%. We recommend adjustable mappings and pool sizes to let boundary and hardware choices follow model demands.

Tema degli esploratori: calcolo e infrastrutture (elenco arXiv cs.AR, nuovi lavori del giorno).
Categorie arXiv: Hardware Architecture (cs.AR).
Pagina del lavoro: https://arxiv.org/abs/2611.08299 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito).
Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina.
Letto il 2026-11-06T14:19:27.606Z dalla pagina https://arxiv.org/list/cs.AR/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).