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
U.S. Stocks Hit Historic Milestone as AI and Aerospace Giants Prepare for Trillion-Dollar Debut
Elon Musk, Sam Altman, and Dario Amodei, three titans of the technology sector, are advancing toward initial public offerings for their respective ventures. With SpaceX, OpenAI, and Anthropic—three industry behemoths nearing trillion-dollar valuation
Swedish AI Startup Lovable Eyes $13.2 Billion Valuation After Major Funding Round
As AI-driven coding tools gain traction, Swedish startup Lovable has secured a major funding round. The company aims to raise $3 billion, potentially boosting its valuation to $13.2 billion—double the $6.6 billion recorded last December. Menlo Ventur
Google Tests Remy AI Agent for Gemini as Focus Shifts to User Control
According to Business Insider, Google is testing Remy, a new AI personal agent for Gemini. This tool aims to execute tasks on behalf of users, streamlining both professional workflows and daily routines.Currently, Remy is undergoing testing in an int
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.
U.S. Stocks Hit Historic Milestone as AI and Aerospace Giants Prepare for Trillion-Dollar Debut
Elon Musk, Sam Altman, and Dario Amodei, three titans of the technology sector, are advancing toward initial public offerings for their respective ventures. With SpaceX, OpenAI, and Anthropic—three industry behemoths nearing trillion-dollar valuation
Swedish AI Startup Lovable Eyes $13.2 Billion Valuation After Major Funding Round
As AI-driven coding tools gain traction, Swedish startup Lovable has secured a major funding round. The company aims to raise $3 billion, potentially boosting its valuation to $13.2 billion—double the $6.6 billion recorded last December. Menlo Ventur





Home






