Energy Consumption Analysis of Discrete Diffusion and Autoregressive Language Models
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).
ProofPaper ↗
Key points
- 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.
Sources (1)
- [1]Energy Consumption Analysis of Discrete Diffusion and Autoregressive Language ModelsarXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 14, 09:21 AM
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).
As Large Language Models (LLMs) continue to scale, their energy consumption during inference has become a critical concern.
Extractive summary: sentences quoted from the sources.
Before this
- Oct 14, 2026PN-Mix: Pair-Norm Guided Mixed-Precision Key Cache Quantization for Efficient Long-Output LLM Inference
- Oct 14, 2026The Design Space of Tri-Modal Masked Diffusion Models
- Oct 13, 2026A Few More Tokens Ahead: Reinforcing Parallelism in Diffusion LLMs
- Oct 13, 2026HaloBlock: Faster and Better Decoding for Diffusion Large Language Models
- Oct 13, 2026Towards Green AI: Evaluating Energy Efficiency for Quantized LLM Inference
- Oct 12, 2026JustOPD: On-Policy Distillation from Autoregressive to Diffusion Language Models
