analiza aiqbench
This workshop paper measures energy per token for two diffusion language models against two autoregressive models of similar size, with and without KV-cache. The reported roughly 2x efficiency edge for diffusion at 200-600 tokens is a useful data point, but it depends on the specific models, sequence lengths, and hardware setup, so it should not be read as a general result. Readers evaluating inference costs should look for replication across hardware and serving stacks before treating diffusion decoding as a settled energy advantage.
Energy Consumption Analysis of Discrete Diffusion and Autoregressive Language Models
arXiv:2610.19855v1 Announce Type: new Abstract: As Large Language Models (LLMs) continue to scale, their energy consumption during inference has become a critical concern. While Autoregressive (AR) models are the industry standard, Diffusion Language Models (DLLMs) have emerged as a promising alternative capable of parallel token generation. In this paper, we perform a comparative energy benchmark of diffusion models (LLaDA-8B, Dream-7B) against AR counterparts (Llama-3.1-8B, Qwen-2.5-7B). We analyze the trade-offs between model architecture, KV-cache optimization, and Joules-per-token efficiency. Our results show that at longer sequences (200--600 tokens), diffusion models achieve 2.5--2.7 J/token, approximately 2$\times$ more efficient than AR models with KV-cache (5.3--5.7 J/token). KV-cache provides substantial benefits for AR models at longer sequences (50--70\% energy reduction).
Abstract:As Large Language Models (LLMs) continue to scale, their energy consumption during inference has become a critical concern. While Autoregressive (AR) models are the industry standard, Diffusion Language Models (DLLMs) have emerged as a promising alternative capable of parallel token generation. In this paper, we perform a comparative energy benchmark of diffusion models (LLaDA-8B, Dream-7B) against AR counterparts (Llama-3.1-8B, Qwen-2.5-7B). We analyze the trade-offs between model architecture, KV-cache optimization, and Joules-per-token efficiency. Our results show that at longer sequences (200--600 tokens), diffusion models achieve 2.5--2.7 J/token, approximately 2$\times$ more efficient than AR models with KV-cache (5.3--5.7 J/token). KV-cache provides substantial benefits for AR models at longer sequences (50--70\% energy reduction).
Machine Learning (cs.LG)
arXiv:2610.19855 [cs.LG]
(or arXiv:2610.19855v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2610.19855 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
GreenSys '26: Proceedings of the 2nd International Workshop on Systems and Methods for Sustainable Large-Scale AI, 2026, pp. 7-12
https://doi.org/10.1145/3802973.3804458 Focus to learn more DOI(s) linking to related resources
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 233 cuvinte
Subiecte conexe
Self-Orchestrating Language Models: Leveraging Semantic Dependence for Efficient Inference
arXiv · 16 sept. 2026
Neuromorphic Diffusion Language Models: Addressing Compute and Memory Bottlenecks via Sparsity and Block Denoising
arXiv · 29 iul. 2026
Google AI Releases DiffusionGemma, a 26B MoE Open Model Using Text Diffusion for Up to 4x Faster Generation
MarkTechPost · 10 iun. 2026
The Design Space of Tri-Modal Masked Diffusion Models
Apple · 14 oct. 2026