JD Ads and Huawei Kunpeng jointly announced that, leveraging the Lingqu (UnifiedBus) interconnect technology, they have compressed the traditional four-level cache into a two-level architecture, built a 300 TB shared memory pool, and rearchitected the RPC communication framework, reducing end-to-end inference latency in their generative advertising recommendation system by 15%. This achievement has been continuously validated across 5% of online traffic in JD.com’s core advertising scenarios during the 2026 618 shopping festival, with system stability, throughput, and low-latency performance all meeting expected targets.
The technological breakthroughs are concentrated in three areas: First, a 4K-node ScaleUp networking scheme based on the Lingqu protocol has been implemented, with each super node supporting over 1.5 million vCPUs and 3.2 Tb of interconnect bandwidth, thereby resolving the cross-layer bandwidth convergence bottleneck. Second, the KV cache adopts a two-tier architecture—on-chip memory plus DDR—reducing inter-cache latency by three orders of magnitude; for 8 MB concurrent requests, TP99 latency drops by more than 85%, and at a sequence length of 64 MB, TP99 latency is reduced to one-sixteenth of its original value. Third, RPC communication bypasses the Socket/TCP/IP protocol stack, switching instead to direct URMA connections, resulting in an 84% reduction in latency for 1 MB data packets, an overall P99 optimization exceeding 40%, and a 100-fold reduction in memory usage for communication.
This collaboration has expanded from advertising into search, recommendation, and other search‑recommendation scenarios, while also launching explorations in new areas such as SuperCache (L2/L3 capacity expansion and unified interconnection of general‑purpose and AI computing), the Agent high‑density sandbox (supporting hundreds of thousands of concurrent connections within seconds, with fivefold resource overcommitment), and multimodal fusion retrieval. Both parties note that in the Agentic AI era, reliance on CPUs will increase significantly, with the CPU‑to‑NPU ratio likely shifting from 1:4 to a higher proportion, necessitating a coordinated rearchitecting of the underlying compute infrastructure—spanning the operating system, scheduling, and middleware.
The report explicitly credits its source as the Titanium Media App (https://www.tmtpost.com/8181692.html），发布于2026-10-26T12:19:13.000Z，属时效性强、细节翔实的技术进展类新闻。文中所有技术指标、验证方式、演进路径均严格依据材料原文提取，未引入外部知识、未推测未声明的商业影响或未发生的未来事件。).