cs.LGNov 8, 2026

Direct GPU Energy Measurements of Batch-Level Sharing in Mixture-of-Experts and Recurrent Models

Authors: Rafael Pina, Varuna De Silva

Organizations: Institute for Digital Technologies, Loughborough University London, London, United Kingdom · Center for Neuromorphic Intelligence, Aston University, Birmingham, United Kingdom

Abstract

Energy claims for efficient neural networks are often inferred from operation counts. We report direct GPU energy measurements on two workloads that batch very differently, recurrent policies in GPU-vectorised multi-agent reinforcement learning (MARL) and mixture-of-experts (MoE) language models, and ask one question of both: can the rows of a batch be made to share computation, across the batch and over time, at a measured saving? On recurrent policies the answer so far is no. Shared activity patterns exist but rarely last one step, masking a frozen policy fails, and trained-in masks succeed only because a fixed mask is equivalent to a narrow GRU, at 11x the energy of a small dense policy in our implementation. Test-time steering adds nothing beyond re-picking on either workload. On MoE models the answer is yes, and granularity decides how much: on Qwen3-30B-A3B, batch-aware routing reads 25-31% fewer expert weights within KL 0.05, with benchmark changes within noise at 500 items, and uses 12-18% less energy per token on a fused forward. We also introduce persistence-aware OEA, a one-parameter rule that keeps the batch-aware cut with about 40% fewer new experts per step than native routing at batch 32-64, though its energy has not yet been measured. These are pilot measurements on one GPU type, and we state their limits.

Explore similar work

Oct 27, 2026

LRMoE: Demand-Side Load Re-Calibration for Efficient MoE Training on Memory-Constrained Nodes

Sparse Mixture-of-Experts (MoE) models offer scalable capacity for Web applications, but uneven routing creates computation and communication bottlenecks. Existing placement and replication methods redistribute this workload, yet additional expert copies compete with token-related buffers for limited GPU memory. We present LRMoE, a memory-aware MoE training system that jointly recalibrates backward demand and balances the retained workload. During forward execution, it plans retention quotas and expert placement under resource constraints while preserving all routed computations. During backpropagation, a load recalibration metric estimates each token-expert pair's contribution to the expert update and selects pairs within these quotas. A capacity-aware routing algorithm distributes retained pairs among eligible instances of their original experts. Physical compaction then forms smaller communication payloads and variable-size grouped GEMMs, reducing backward traffic, computation, and execution-buffer demand. Experiments on Qwen3-30B-A3B and GLM4.5-106B-A12B show that LRMoE reduces cumulative MoE-layer training time to fixed-loss targets by 21.2-44.8% relative to SYMI.
Nov 2, 2026

Sonar-MoE: In-Backward Expert Updates for Memory-Efficient Mixture-of-Experts Training

Most parameters of a Mixture-of-Experts (MoE) model belong to its experts, and standard training keeps every expert gradient in GPU memory until the optimizer step. This memory grows with the number and size of the experts, so it rises as MoE models scale. We introduce Sonar-MoE, which updates each expert tensor during the backward pass and frees its gradient right away. It brings optimizer-in-backward from dense models to batched MoE expert tensors, so the GPU never holds more than one layer of expert gradients. When we fine-tune OLMoE-1B-7B and Granite-3.1-3B-A800M in bf16 with two expert-kernel backends, Sonar-MoE lowers peak memory by 19% to 22%. Its updates are bitwise identical to standard training without expert gradient clipping, and held-out loss stays within 0.003 of all baselines. Each step also takes 17% to 23% less time than standard training with global-norm clipping. Updating whole batched tensors rather than individual experts is what keeps Sonar-MoE fast, as it avoids a 20% to 72% slowdown while using the same memory. Sonar-MoE matches optimizer-in-backward over all parameters in memory and speed while still clipping the dense parameters.
Apr 25, 2026

Scaling Multi-Node Mixture-of-Experts Inference Using Expert Activation Patterns

Most recent state-of-the-art (SOTA) large language models (LLMs) use Mixture-of-Experts (MoE) architectures to scale model capacity without proportional per-token compute, enabling higher-quality outputs at manageable serving costs. However, MoE inference at scale is fundamentally bottlenecked by expert load imbalance and inefficient token routing, especially in multi-node deployments where tokens are not guaranteed to be routed to local experts, resulting in significant inter-node all-to-all communication overhead. To systematically characterize these challenges, we profile SOTA open-source MoE models, including Llama 4 Maverick, DeepSeek V3-671B, and Qwen3-230B-A22B, on various datasets and collected over 100k real expert activation traces. Upon studying the expert activation patterns, we uncover various persistent properties across all the frontier MoE models: variable expert load imbalance, domain-specific expert activation where expert popularity shifts across task families (code, math, chat, general), and a strong correlation between prefill and decode expert activations. Motivated by these findings, we propose workload-aware micro-batch grouping and an expert placement strategy to maximize token locality to the destination expert, thereby reducing inter-node communication. Across models and datasets, these optimizations help reduce all2all communication data up to 20, resulting in lower MoE decode latency and better accelerator utilization.