SQD disaggregates decode by quadratic and subquadratic attention stages instead of treating decode as one pool or splitting only attention from feed-forward work. On an adjusted eight-B200 proxy, it improves tokens per joule by 31–56% over the strongest GPU-only baselines across GLM-5.2, Nemotron 3 Ultra and Gemma 4 31B. A validated analytical model of Rubin plus LPX projects 1.2–1.5 times tighter achievable latency and up to 3.6 times the throughput of attention-FFN disaggregation. The proxy modifies decode-device characteristics and most future-system results are modeled, not production measurements.
SQD splits decode around subquadratic attention stages
2026-09-14Frontier watchWATCH
SQD disaggregates decode by quadratic and subquadratic attention stages instead of treating decode as one pool or splitting only attention from feed-forward work. On an adjusted eight-B200 proxy, it improves tokens per joule by 31–56% over the strongest GPU-only baselines across GLM-5.2, Nemotron 3 Ultra and Gemma 4 31B. A validated analytical model of Rubin plus LPX projects 1.2–1.5 times tighter achievable latency and up to 3.6 times the throughput of attention-FFN disaggregation. The proxy modifies decode-device characteristics and most future-system results are modeled, not production measurements.
Availability: paper and methodology only; no implementation repository linked
Imported from the existing research archive. Migration preserves the recorded judgment and is not a fresh verification.
What this means for you
For inference architects planning mixed GPU and decode-accelerator fleets, hybrid and sparse attention change the correct disaggregation boundary. Moving fixed-state attention with the FFN can reduce energy and interconnect pressure. The paper is architecture guidance rather than deployable software, so capacity plans should wait for a real Rubin/LPX implementation and end-to-end validation.
FRONTIER WATCH
Public artifacts and validated results are different.
The recorded fact is the public work and its artifacts. Author-reported outcomes do not establish independent reproduction. Next checks include usable code, task conditions and third-party validation.
Evidence & related material
Technical details & scores
| Models | GLM-5.2 sparse attention; Nemotron 3 Ultra linear/recurrent attention; Gemma 4 31B sliding-window attention |
|---|---|
| Proxy | adjusted 8xB200 system with modified decode-device bandwidth, capacity and power assumptions |
| Measured Energy | 31-56% average tokens/J gain over strongest GPU-only baselines; 53-71% at 1M context |
| Modeled Future System | Rubin plus LPX, 1.2-1.5x tighter achievable latency and up to 3.6x throughput over attention-FFN disaggregation |
| Cost | 4,058 B200 GPU-hours across 579 runs; 358 hours contributed reported output |
| Availability | paper and methodology only; no implementation repository linked |
Technical impact 4
Engineering value 4
Adoption signal 1
Maturity 1
Verification cost 5
Risk 3