Tsinghua and Tencent Hunyuan Win MLSys2026 MoE Inference Challenge with 4.1x NPU Speedup
Tsinghua University's Storage Lab and Tencent's MegEngine AI Infra team recently claimed the global championship in the MoE Model Inference Optimization Challenge at MLSys2026, a leading machine learning systems conference.

To tackle inference performance bottlenecks in the mixture-of-experts (MoE) architecture with trillion-scale parameters running on heterogeneous NPU chips, the joint team designed a full-chain optimization solution for the official model and NPU hardware. By employing the E-Shard strategy for expert-level task partitioning, PSUM 3D tensor batch readout, the GEMV path distributing outputs across multiple banks for concurrency, and using scalar engines to reduce initial data transfer latency, they successfully resolved low data transfer efficiency and repeated activation transfers at the operator level.
In parallel, the team reorganized the on-chip data layout for the attention module and integrated core Transformer operators, achieving bit-level high-precision alignment.

Figure 3: Diagram of the MoE optimization architecture, featuring E-Shard expert partitioning, continuous DMA, PSUM/GEMV concurrency, cold start pipeline, and prefetch control.
Additionally, the team developed an agent-based inference operator optimizer named "Knight". Through an automated closed-loop process of proposal, code implementation, and iteration, it dramatically broadened the optimization search space. As a result, the end-to-end inference time dropped from 14.91 seconds to 3.56 seconds, a 4.1x speedup; single-step decoding latency fell from 12.63ms to 5.45ms, and DMA engine utilization during weight loading climbed to around 80%.
Outperforming teams from leading global universities like Stanford and MIT, this achievement highlights the deep expertise of Chinese teams in adapting large models to underlying systems and optimizing operators. It also provides a valuable engineering blueprint for deploying trillion-parameter MoE models on future super-node computing platforms.
Related article
Agility Robotics Selects Nyobolt for Rapid Charging of Digit 5 Humanoid
Nyobolt supplies fast-charging, long-lasting battery technology for Digit 5. Source: AgilityPower management remains a critical hurdle for humanoid robot developers. Agility Robotics Inc. announced a collaboration with Nyobolt to develop a rapid-rech
Apple iOS 27 Update: Siri Enters Dual-Mode Era with ChatGPT Free Switch
iOS 27 introduces a fundamental overhaul of Siri’s architecture, centered on a new standalone application that moves beyond traditional voice-only interactions.Designed with a chatbot interface, the app allows users to type prompts for seamless AI co
Soul Founder Unveils Open Core AI Tech as 4,700 Students Redefine Gen Z Socializing at Meta Creation Camp
As artificial intelligence reshapes social media, Gen Z’s approach to meaningful connection is evolving. On February 27, 2026, data from Qichacha revealed that the founder of Soul is actively driving the "2025 Soul App Metaverse Creation Camp College
Related Special Topic Recommendations
Comments (1)
0/500
Tsinghua University's Storage Lab and Tencent's MegEngine AI Infra team recently claimed the global championship in the MoE Model Inference Optimization Challenge at MLSys2026, a leading machine learning systems conference.

To tackle inference performance bottlenecks in the mixture-of-experts (MoE) architecture with trillion-scale parameters running on heterogeneous NPU chips, the joint team designed a full-chain optimization solution for the official model and NPU hardware. By employing the E-Shard strategy for expert-level task partitioning, PSUM 3D tensor batch readout, the GEMV path distributing outputs across multiple banks for concurrency, and using scalar engines to reduce initial data transfer latency, they successfully resolved low data transfer efficiency and repeated activation transfers at the operator level.
In parallel, the team reorganized the on-chip data layout for the attention module and integrated core Transformer operators, achieving bit-level high-precision alignment.

Figure 3: Diagram of the MoE optimization architecture, featuring E-Shard expert partitioning, continuous DMA, PSUM/GEMV concurrency, cold start pipeline, and prefetch control.
Additionally, the team developed an agent-based inference operator optimizer named "Knight". Through an automated closed-loop process of proposal, code implementation, and iteration, it dramatically broadened the optimization search space. As a result, the end-to-end inference time dropped from 14.91 seconds to 3.56 seconds, a 4.1x speedup; single-step decoding latency fell from 12.63ms to 5.45ms, and DMA engine utilization during weight loading climbed to around 80%.
Outperforming teams from leading global universities like Stanford and MIT, this achievement highlights the deep expertise of Chinese teams in adapting large models to underlying systems and optimizing operators. It also provides a valuable engineering blueprint for deploying trillion-parameter MoE models on future super-node computing platforms.
Apple iOS 27 Update: Siri Enters Dual-Mode Era with ChatGPT Free Switch
iOS 27 introduces a fundamental overhaul of Siri’s architecture, centered on a new standalone application that moves beyond traditional voice-only interactions.Designed with a chatbot interface, the app allows users to type prompts for seamless AI co
Soul Founder Unveils Open Core AI Tech as 4,700 Students Redefine Gen Z Socializing at Meta Creation Camp
As artificial intelligence reshapes social media, Gen Z’s approach to meaningful connection is evolving. On February 27, 2026, data from Qichacha revealed that the founder of Soul is actively driving the "2025 Soul App Metaverse Creation Camp College





Home






