# EMA: Elastic and Performance Transparent Memory Across GPUs

arXiv: https://arxiv.org/abs/2609.27040 · HTML: https://mlsanity.com/paper/2609.27040/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 26 of its 55 references that are in the map.
- First posted on arXiv 22 Sep 2026.

## Builds on (most cited first, 26 of 26)

- [PyTorch: An Imperative Style, High-Performance Deep Learning Library](https://mlsanity.com/paper/1912.01703/) (2019)
- [LLaMA: Open and Efficient Foundation Language Models](https://mlsanity.com/paper/2302.13971/) (2023)
- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness](https://mlsanity.com/paper/2205.14135/) (2022)
- [OPT: Open Pre-trained Transformer Language Models](https://mlsanity.com/paper/2205.01068/) (2022)
- [FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning](https://mlsanity.com/paper/2307.08691/) (2023)
- [Mixed Precision Training](https://mlsanity.com/paper/1710.03740/) (2017)
- [Training Deep Nets with Sublinear Memory Cost](https://mlsanity.com/paper/1604.06174/) (2016)
- [FlexGen: High-Throughput Generative Inference of Large Language Models with a Single GPU](https://mlsanity.com/paper/2303.06865/) (2023)
- [DeepSpeed Inference: Enabling Efficient Inference of Transformer Models at Unprecedented Scale](https://mlsanity.com/paper/2207.00032/) (2022)
- [ZeRO-Offload: Democratizing Billion-Scale Model Training](https://mlsanity.com/paper/2101.06840/) (2021)
- [Reducing Activation Recomputation in Large Transformer Models](https://mlsanity.com/paper/2205.05198/) (2022)
- [Clipper: A Low-Latency Online Prediction Serving System](https://mlsanity.com/paper/1612.03079/) (2016)
- [Checkmate: Breaking the Memory Wall with Optimal Tensor Rematerialization](https://mlsanity.com/paper/1910.02653/) (2019)
- [TensorFlow-Serving: Flexible, High-Performance ML Serving](https://mlsanity.com/paper/1712.06139/) (2017)
- [Pollux: Co-adaptive Cluster Scheduling for Goodput-Optimized Deep Learning](https://mlsanity.com/paper/2008.12260/) (2020)
- [TurboTransformers: An Efficient GPU Serving System For Transformer Models](https://mlsanity.com/paper/2010.05680/) (2020)
- [SuperNeurons: Dynamic GPU Memory Management for Training Deep Neural Networks](https://mlsanity.com/paper/1801.04380/) (2018)
- [Evaluating Modern GPU Interconnect: PCIe, NVLink, NV-SLI, NVSwitch and GPUDirect](https://mlsanity.com/paper/1903.04611/) (2019)
- [M'elange: Cost Efficient Large Language Model Serving by Exploiting GPU Heterogeneity](https://mlsanity.com/paper/2404.14527/) (2024)
- [Optimal checkpointing for heterogeneous chains: how to train deep neural networks with limited memory](https://mlsanity.com/paper/1911.13214/) (2019)
- [Demystifying Cost-Efficiency in LLM Serving over Heterogeneous GPUs](https://mlsanity.com/paper/2502.00722/) (2025)
- [AQUA: Network-Accelerated Memory Offloading for LLMs in Scale-Up GPU Domains](https://mlsanity.com/paper/2407.21255/) (2024)
- [Pie: Pooling CPU Memory for LLM Inference](https://mlsanity.com/paper/2411.09317/) (2024)
- [ProTrain: Efficient LLM Training via Memory-Aware Techniques](https://mlsanity.com/paper/2406.08334/) (2024)
- [Harvest: Opportunistic Peer-to-Peer GPU Caching for LLM Inference](https://mlsanity.com/paper/2602.00328/) (2026)

## Abstract (arXiv)

Multi-GPU servers have become the standard building block of modern data centers, providing aggregated capacity through high-bandwidth interconnects. At the same time, workloads such as LLM inference exhibit highly dynamic memory demands, which can cause one GPU to exhaust its local memory while others remain underutilized. This mismatch motivates a model of elastic resource sharing across GPUs. We present EMA, a memory sharing system that allows GPUs within a server to borrow and reclaim memory from each other, forming an elastic pool of capacity. EMA ensures performance transparency for both borrowers and lenders. For borrowers, prefetching hides remote access costs so that applications experience remote and local memory as indistinguishable in performance. For lenders, borrowed resources remain reclaimable on demand, guaranteeing that performance never falls below that of static partitioning. While our design focuses on memory, the same principle naturally extends to other GPU resources. Our evaluation shows that EMA improves individual user throughput by up to 52%, achieves 96% of the throughput of a system provisioned with 2X capacity, and maintains latency similar to the static local baseline.