← Research Lab / Systems
FIG. 0570 // FANTASMA RESEARCH NOTE

Latency Budget Allocation Across a Distributed Inference Pipeline

Oleh Yuki Tanaka
7 min baca
2026-01-22

Latency Budget Allocation

A real-time inference pipeline is a chain of dependent operations. Each has a latency distribution. The end-to-end latency is not the sum of medians — it's a convolution of distributions, and tail behavior dominates.

When your SLA is p95 ≤ 12ms and you have 6 pipeline stages, allocating the latency budget is a non-trivial problem. Naive allocation (2ms per stage) ignores the variance structure. A single high-variance stage can blow the entire budget.

Formalization

We model each stage i as having latency distribution Lᵢ with known mean μᵢ and variance σᵢ². The pipeline latency is:

L_total = Σ Lᵢ + Σ queueing_delays

For a target p95 SLA, we derive per-stage budget allocations that minimize the probability of SLA violation, subject to the constraint that allocations sum to the total budget.

The Scheduling Policy

The optimal allocation assigns tighter budgets to high-variance stages (to reduce tail risk) and looser budgets to low-variance stages. In practice, we implement this via an admission control mechanism: stages that are predicted to exceed their budget shed load to a fallback (lower-quality but faster) path.

Results

Applied to Neural Fabric's inference pipeline: SLA violation rate reduced from 0.8% to 0.04% with no change in median latency.

Publikasi Riset Terkait

Architecture 11 min

Attention Is Not Enough: Toward Structured World Models at the Edge

Transformer attention scales quadratically with sequence length. For always-on edge inference, this is a thermodynamic impossibility. We propose a hybrid SSM-attention architecture that achieves 94% of full-attention quality at 8% of the compute.

HCI 8 min

Calibration-Free Gaze Tracking at 60fps on Commodity Hardware

Classical gaze tracking requires a 90-second calibration ritual that most users abandon. We trained a universal gaze model on 4.2M synthetic eye-renders that generalizes to unseen users with 0.4° accuracy — no calibration, no personal data.