Home
Moonshot AI and Tsinghua University Unveil PrfaaS Architecture to Break Large Model Computing Power Bottleneck
New technologies are breaking the performance bottleneck of large language models (LLMs). Recently, researchers from Moonshot AI (Moonshot) and Tsinghua University introduced a novel architecture called Prefill-as-a-Service (PrfaaS). This study tackles hardware constraints in data center deployments of large model services by optimizing compute resource allocation, thereby significantly boosting inference efficiency.

Technical Breakthrough: A "Surgical" Separation of Prefill and Decode
Currently, the inference process for large language models comprises two distinct phases:
Prefill Phase: This phase is compute-intensive, handling input processing and generating key-value cache (KVCache).
Decode Phase: This phase is memory-bandwidth-intensive, generating output word by word.
In traditional architectures, both phases are typically handled within the same data center or even the same server. Because they have different hardware resource needs, this "forced bundling" often creates an imbalance in compute and bandwidth allocation, leading to service congestion.
Core Innovation: Efficient Cross-Region Collaboration
PrfaaS ’s core highlight is achieving a decoupled service. It offloads the compute-intensive prefill tasks to specialized high-compute clusters. After task completion, the system uses standard Ethernet to transfer the generated KVCache remotely to a local decoding cluster.
This design removes physical space constraints, enabling prefill and decoding to run concurrently across different data centers. To ensure efficient transfer, PrfaaS incorporates a two-time-scale scheduling mechanism that flexibly allocates resources based on real-time traffic fluctuations, paired with precise routing to prevent long text requests from being delayed by uneven resource distribution.
Test Results: Optimizing Throughput and Latency
Research data demonstrates that the PrfaaS architecture delivers remarkable performance in real-world applications:
Service throughput increased by 54%, significantly improving the number of requests handled per unit time.
Response latency decreased significantly, making the first token generation faster from the user's perspective.
Maximized resource utilization by separating compute, network, and storage subsystems, avoiding congestion issues seen in traditional architectures.
The collaboration between Moonshot AI and Tsinghua University not only offers new engineering approaches for large-scale AI inference but also lays the groundwork for future cross-regional computing networks. This Prefill-as-a-Service model could become a key milestone in moving large models toward industrial deployment.
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
New technologies are breaking the performance bottleneck of large language models (LLMs). Recently, researchers from Moonshot AI (Moonshot) and Tsinghua University introduced a novel architecture called Prefill-as-a-Service (PrfaaS). This study tackles hardware constraints in data center deployments of large model services by optimizing compute resource allocation, thereby significantly boosting inference efficiency.

Technical Breakthrough: A "Surgical" Separation of Prefill and Decode
Currently, the inference process for large language models comprises two distinct phases:
Prefill Phase: This phase is compute-intensive, handling input processing and generating key-value cache (KVCache).
Decode Phase: This phase is memory-bandwidth-intensive, generating output word by word.
In traditional architectures, both phases are typically handled within the same data center or even the same server. Because they have different hardware resource needs, this "forced bundling" often creates an imbalance in compute and bandwidth allocation, leading to service congestion.
Core Innovation: Efficient Cross-Region Collaboration
This design removes physical space constraints, enabling prefill and decoding to run concurrently across different data centers. To ensure efficient transfer, PrfaaS incorporates a two-time-scale scheduling mechanism that flexibly allocates resources based on real-time traffic fluctuations, paired with precise routing to prevent long text requests from being delayed by uneven resource distribution.
Test Results: Optimizing Throughput and Latency
Research data demonstrates that the PrfaaS architecture delivers remarkable performance in real-world applications:
Service throughput increased by 54%, significantly improving the number of requests handled per unit time.
Response latency decreased significantly, making the first token generation faster from the user's perspective.
Maximized resource utilization by separating compute, network, and storage subsystems, avoiding congestion issues seen in traditional architectures.
The collaboration between Moonshot AI and Tsinghua University not only offers new engineering approaches for large-scale AI inference but also lays the groundwork for future cross-regional computing networks. This Prefill-as-a-Service model could become a key milestone in moving large models toward industrial deployment.
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











