As GPU capacity fragments across geographically distributed sites, single-cluster LLM inference routing assumptions break down in measurable ways. We present Solyx AI Grid, a cross-site inference routing control plane that integrates GPU hardware telemetry (DCGM), vLLM application metrics, and real-time WAN signals (RTT, jitter) into per-request placement decisions via a 10-signal weighted pressure scorer. Across two empirical campaigns--six H100/H200 SXM GPUs and nine RTX PRO 6000 Blackwell SE GPUs spanning three US datacenters, eight workload classes, and a 216-cell SLO matrix--Solyx AI Grid delivers 1.56--1.75x throughput at tier-2 SLO over round-robin across all eight classes, cuts capability-mismatch leakage to 0.43% (versus 32% for standard routers), and reroutes around failures at a p99 of 1,247 ms versus 4,226 ms. We further find that GPU hardware telemetry leads application-layer SLO breach by 11.2 seconds on average, enabling proactive traffic drain before user-facing latency impact. To our knowledge, this is the first public empirical study of live physical multi-site LLM inference routing combining hardware telemetry, application metrics, and active WAN path signals.

翻译：随着GPU容量分散于地理分布式站点，单集群大语言模型推理路由假设在可量化维度上失效。我们提出Solyx AI Grid——一种跨站点推理路由控制平面，通过集成GPU硬件遥测（DCGM）、vLLM应用指标及实时广域网信号（RTT、抖动），基于10信号加权压力评分器实现每请求放置决策。在两轮实证研究（涵盖三个美国数据中心、八类负载及216单元SLO矩阵的六块H100/H200 SXM GPU与九块RTX PRO 6000 Blackwell SE GPU）中，Solyx AI Grid在全部八类负载的第二级SLO约束下实现1.56至1.75倍于轮询的吞吐量，将能力失配泄漏降至0.43%（标准路由器为32%），并以p99 1,247毫秒（对比4,226毫秒）完成故障重路由。进一步发现，GPU硬件遥测信号平均比应用层SLO违例提前11.2秒触发，使得在用户感知延迟影响前即可实现主动流量疏散。据我们所知，这是首个结合硬件遥测、应用指标与主动广域网路径信号的实时物理多站点LLM推理路由公开实证研究。