# Rail-only: A Low-Cost High-Performance Network for Training LLMs with Trillion Parameters

arXiv: https://arxiv.org/abs/2307.12169 · HTML: https://mlsanity.com/paper/2307.12169/

## Key facts

- Cited by 13 papers in the map (data through 14 Oct 2026).
- Placed in Efficient MoE training and inference systems, one of 12 directions in Mixture-of-experts.
- Builds on 12 of its 45 references that are in the map.
- First posted on arXiv 22 Jul 2023.

## Placed in

- Mixture-of-experts / [Efficient MoE training and inference systems](https://mlsanity.com/moe/efficient-moe-training-and-inference-systems/)

## Builds on (most cited first, 12 of 12)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Scaling Laws for Neural Language Models](https://mlsanity.com/paper/2001.08361/) (2020)
- [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://mlsanity.com/paper/1909.08053/) (2019)
- [Efficient Large-Scale Language Model Training on GPU Clusters Using Megatron-LM](https://mlsanity.com/paper/2104.04473/) (2021)
- [Efficiently Scaling Transformer Inference](https://mlsanity.com/paper/2211.05102/) (2022)
- [DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale](https://mlsanity.com/paper/2201.05596/) (2022)
- [Reducing Activation Recomputation in Large Transformer Models](https://mlsanity.com/paper/2205.05198/) (2022)
- [TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings](https://mlsanity.com/paper/2304.01433/) (2023)
- [Software-Hardware Co-design for Fast and Scalable Training of Deep Learning Recommendation Models](https://mlsanity.com/paper/2104.05158/) (2021)
- [TopoOpt: Co-optimizing Network Topology and Parallelization Strategy for Distributed Training Jobs](https://mlsanity.com/paper/2202.00433/) (2022)
- [Expanding across time to deliver bandwidth efficiency and low latency](https://mlsanity.com/paper/1903.12307/) (2019)
- [Bandwidth Optimal Pipeline Schedule for Collective Communication](https://mlsanity.com/paper/2305.18461/) (2023)

## Built on by (most cited first, 13 of 13)

- [Revisiting Reliability in Large-Scale Machine Learning Research Clusters](https://mlsanity.com/paper/2410.21680/) (2024)
- [KVDirect: Distributed Disaggregated LLM Inference](https://mlsanity.com/paper/2501.14743/) (2024)
- [InfiniteHBD: Building Datacenter-Scale High-Bandwidth Domain for LLM with Optical Circuit Switching Transceivers](https://mlsanity.com/paper/2502.03885/) (2025)
- [Training Foundation Models on a Full-Stack AMD Platform: Compute, Networking, and System Design](https://mlsanity.com/paper/2511.17127/) (2025)
- [ForestColl: Throughput-Optimal Collective Communications on Heterogeneous Network Fabrics](https://mlsanity.com/paper/2402.06787/) (2024)
- [Exploring Scaling Laws for Local SGD in Large Language Model Training](https://mlsanity.com/paper/2409.13198/) (2024)
- [Scalable Synthesis of distributed LLM workloads through Symbolic Tensor Graphs](https://mlsanity.com/paper/2511.10480/) (2025)
- [Folding Tensor and Sequence Parallelism for Memory-Efficient Transformer Training & Inference](https://mlsanity.com/paper/2604.26294/) (2026)
- [Rethinking Network Topologies for Cost-Effective Mixture-of-Experts LLM Serving](https://mlsanity.com/paper/2605.00254/) (2026)
- [Resilient AI Supercomputer Networking using MRC and SRv6](https://mlsanity.com/paper/2605.04333/) (2026)
- [Diagnosing Overhead in Dispatch Operations: Cross-architecture Observatory](https://mlsanity.com/paper/2605.20982/) (2026)
- [mKernel: Fast Multi-GPU, Multi-Node Fused Kernels](https://mlsanity.com/paper/2609.13585/) (2026)
- [The Life of a Token: from Words to Bits on the Wire](https://mlsanity.com/paper/2609.19924/) (2026)

## Abstract (arXiv)

This paper presents a low-cost network architecture for training large language models (LLMs) at hyperscale. We study the optimal parallelization strategy of LLMs and propose a novel datacenter network design tailored to LLM's unique communication pattern. We show that LLM training generates sparse communication patterns in the network and, therefore, does not require any-to-any full-bisection network to complete efficiently. As a result, our design eliminates the spine layer in traditional GPU clusters. We name this design a Rail-only network and demonstrate that it achieves the same training performance while reducing the network cost by 38% to 77% and network power consumption by 37% to 75% compared to a conventional GPU datacenter. Our architecture also supports Mixture-of-Expert (MoE) models with all-to-all communication through forwarding, with only 8.2% to 11.2% completion time overhead for all-to-all traffic. We study the failure robustness of Rail-only networks and provide insights into the performance impact of different network and training parameters.