NVIDIA shows DSX MaxLPS tech, AI factory token throughput up to 40% per MW
英伟达展示 DSX MaxLPS 技术，AI 工厂每兆瓦 token 吞吐量最高提升 40%
For users managing AI data centers, this technology offers a significant efficiency boost, potentially reducing operational costs by optimizing power usage.
At the AI infrastructure summit, NVIDIA showcased its DSX MaxLPS power scheduling system for AI factories, which boosts token throughput by up to 40% per megawatt.
The system, integrated by Emerald AI into the Conductor platform, dynamically adjusts data center loads based on grid conditions.
Lambda, a GPU cloud service provider, tested the platform and saw a 24% increase in token throughput under fixed power budgets.
In a test with 19 nodes, throughput rose from about 4 million tokens per second to 5 million, a 24% increase and a 23% improvement in per-watt performance.
NVIDIA also disclosed that the Groq 3 LPX rack can add up to 40% more GPUs within the same power budget, increasing token throughput by up to 35% without new power lines.
In a 100K context, Qwen 3.8 27B workload, Groq 3 LPX achieved 2,529 output tokens per user per second.
英伟达展示 DSX MaxLPS 技术，AI 工厂每兆瓦 token 吞吐量最高提升 40%
IT之家 9 月 17 日消息，科技媒体 Wccftech 昨日（9 月 16 日）发布博文，报道称在 AI 基础设施峰会上，英伟达展示了面向 AI 工厂的 DSX MaxLPS 功耗调度系统， 每兆瓦 token 吞吐量最高提升 40%。 IT之家注：每兆瓦 token 吞吐量是指数据中心每消耗 1 MW 电力时，AI 系统每秒能够生成或处理多少个 token。该指标把功耗纳入计算，因此更适合衡量 AI 数据中心在电力受限条件下的产出能力。 英伟达 DSX MaxLPS 技术已由 Emerald AI 集成至 Conductor 平台，用于电网条件下的数据中心负载编排。 Conductor AI 平台中的 DSX Flex 可依据电网状态动态调整数据中心用电需求。平台已接收 200 次电网状态信号，相关操作均未要求用户干预。Lambda 等云服务商测试后，在固定功率预算下取得 24% 的 token 吞吐量提升。 Lambda 是一家 GPU 云服务商，客户数量超过 10,000 家。测试集群包含 5 个机架和 19 个节点。在与 16 个节点满功率运行相同的功率预算下，19 个节点的集群吞吐量从约每秒 400 万个 token 提升至每秒 500 万个，增幅为 24%，每瓦性能提升 23%。 NVIDIA 还披露，Vera Rubin 和 Groq 3 LPX 机架在相同站点功耗范围内，GPU 数量最多可增加 40%。在不新增电力线路的条件下，token 吞吐量最多可提升 35%。在 100K 上下文、Qwen 3.8 27B 工作负载中，Groq 3 LPX 达到每位用户每秒 2,529 个输出 token。 多家人工智能企业已参与早期测试 Vera CPU，相关结果显示性能最高提升 1.9 倍、吞吐量提升 73%、查询任务吞吐量提升 3.3 倍。 Perplexity 测得 SPACE 安全沙箱启动速度提升 1.9 倍，DeepInfra 测得编排步骤延迟速度提升 2.2 倍。 Redpanda 的测试显示，Vera CPU 相比其他 CPU 延迟降低 5.5 倍，吞吐量提升 73%。Starburst 测得查询吞吐量提升 3 倍，Kinetica 测得分析查询性能提升 2.7 倍。
This report is categorized under AITOP's "Models" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to US-China foundation model benchmarks and commercialization comparison. Key entities and technical topics under continuous tracking include NVIDIA, AI infrastructure, token throughput, power efficiency.
Source attribution: curated from IT之家. Cross-source corroboration: 11 independent sources have covered this development (including NVIDIA AI, shao__meng, techcrunch), confirming industry-wide signal strength.
