Moonshot AI and Tsinghua University Propose PrfaaS Architecture for Data Centers
As large language models (LLMs) demand more computational power during reasoning, conventional service architectures encounter bottlenecks. Moonshot AI, in collaboration with a research team from Tsinghua University, has introduced a novel architecture called Pre-filling as a Service (PrfaaS), aimed at overcoming the constraints of data centers and computing resources in LLM services.

The reasoning process of large language models is generally split into two phases: pre-filling and decoding. Pre-filling is computationally intensive, where the model processes input and creates a key-value cache (KVCache). Decoding is memory bandwidth-intensive, generating outputs sequentially. Traditional setups require both stages to occur within the same data center, leading to constraints in computation and bandwidth.
PrfaaS enables efficient cross-data-center service by offloading pre-filling tasks to dedicated high-compute clusters, using a standard Ethernet network to transfer the generated KVCache to local decoding clusters. Research indicates that this architecture boosts processing performance, achieving a 54% increase in service throughput over traditional models. Real-world case studies also show lower latency and higher efficiency.
The PrfaaS architecture decouples the three main subsystems—computing, networking, and storage—managing each independently. A precise routing mechanism ensures efficient transmission of long requests, preventing congestion from uneven resource allocation seen in traditional approaches. Additionally, a dual-timescale scheduling mechanism adapts to varying traffic patterns, further optimizing resource usage.
As the demand for cross-data-center reasoning grows and new hardware continues to emerge, PrfaaS offers a promising solution for future AI applications.
Related article
Nvidia H200 China deal survives Trump-Xi summit, but not as expected
President Trump traveled to Beijing with Jensen Huang joining the delegation at the eleventh hour, departing two days later with vague remarks about potential chip export developments. No deal materialized. Since Trump first authorized sales in Decem
How to fix Core Web Vitals for SEO ranking?
Creating visually appealing assets is crucial in game development. Pixel art, with its charming retro aesthetic, remains a popular choice. With the advent of AI tools, generating pixel art has become more accessible than ever. This guide explores how
TechCrunch Disrupt 2026 breakout session agenda unveiled
TechCrunch Disrupt 2026 arrives at Moscone West in San Francisco, October 13–15, uniting 10,000+ founders, investors, operators, and tech leaders for three days of insight, networking, and forward-looking dialogue. Register now to save up to $100 on
Related Special Topic Recommendations
Comments (10)
0/500
Интересная архитектура PrfaS от Moonshot и Цинхуа для решения проблем с вычислительной мощностью LLM. Надеюсь, это поможет оптимизировать центры обработки данных! 🚀
A arquitetura PrfaS proposta por Moonshot e Tsinghua é um passo interessante para resolver gargalos de computação em LLMs. Será que isso viabiliza data centers mais eficientes? 🤔
La arquitectura PrfaS de Moonshot y Tsinghua parece una gran solución para los cuellos de botella en centros de datos con LLMs. ¡Esperamos ver resultados concretos pronto! 📈
Moonshot AI und Tsinghua zeigen mit PrfaS einen vielversprechenden Ansatz gegen Rechenengpässe bei LLMs. Spannend zu sehen, wie diese Kooperation die Datenzukunft gestaltet! 🌐
As large language models (LLMs) demand more computational power during reasoning, conventional service architectures encounter bottlenecks. Moonshot AI, in collaboration with a research team from Tsinghua University, has introduced a novel architecture called Pre-filling as a Service (PrfaaS), aimed at overcoming the constraints of data centers and computing resources in LLM services.

The reasoning process of large language models is generally split into two phases: pre-filling and decoding. Pre-filling is computationally intensive, where the model processes input and creates a key-value cache (KVCache). Decoding is memory bandwidth-intensive, generating outputs sequentially. Traditional setups require both stages to occur within the same data center, leading to constraints in computation and bandwidth.
PrfaaS enables efficient cross-data-center service by offloading pre-filling tasks to dedicated high-compute clusters, using a standard Ethernet network to transfer the generated KVCache to local decoding clusters. Research indicates that this architecture boosts processing performance, achieving a 54% increase in service throughput over traditional models. Real-world case studies also show lower latency and higher efficiency.
The PrfaaS architecture decouples the three main subsystems—computing, networking, and storage—managing each independently. A precise routing mechanism ensures efficient transmission of long requests, preventing congestion from uneven resource allocation seen in traditional approaches. Additionally, a dual-timescale scheduling mechanism adapts to varying traffic patterns, further optimizing resource usage.
As the demand for cross-data-center reasoning grows and new hardware continues to emerge, PrfaaS offers a promising solution for future AI applications.
Nvidia H200 China deal survives Trump-Xi summit, but not as expected
President Trump traveled to Beijing with Jensen Huang joining the delegation at the eleventh hour, departing two days later with vague remarks about potential chip export developments. No deal materialized. Since Trump first authorized sales in Decem
How to fix Core Web Vitals for SEO ranking?
Creating visually appealing assets is crucial in game development. Pixel art, with its charming retro aesthetic, remains a popular choice. With the advent of AI tools, generating pixel art has become more accessible than ever. This guide explores how
TechCrunch Disrupt 2026 breakout session agenda unveiled
TechCrunch Disrupt 2026 arrives at Moscone West in San Francisco, October 13–15, uniting 10,000+ founders, investors, operators, and tech leaders for three days of insight, networking, and forward-looking dialogue. Register now to save up to $100 on
Интересная архитектура PrfaS от Moonshot и Цинхуа для решения проблем с вычислительной мощностью LLM. Надеюсь, это поможет оптимизировать центры обработки данных! 🚀
A arquitetura PrfaS proposta por Moonshot e Tsinghua é um passo interessante para resolver gargalos de computação em LLMs. Será que isso viabiliza data centers mais eficientes? 🤔
La arquitectura PrfaS de Moonshot y Tsinghua parece una gran solución para los cuellos de botella en centros de datos con LLMs. ¡Esperamos ver resultados concretos pronto! 📈
Moonshot AI und Tsinghua zeigen mit PrfaS einen vielversprechenden Ansatz gegen Rechenengpässe bei LLMs. Spannend zu sehen, wie diese Kooperation die Datenzukunft gestaltet! 🌐





Home






