GPT-OSS-20B用专家混合模型实现更低显存和能耗，适合部署优化。
GPT-OSS-20B: A Comprehensive Deployment-Centric Analysis of OpenAI's Open-Weight Mixture of Experts Model
- 采用专家混合架构，仅激活17.3%参数（3.61B/20.9B）
- 在2048上下文、64解码长度下，吞吐量高31.8%，能耗低25.8%
- 峰值显存减少31.7%，每活跃参数效率显著提升
我们对GPT-OSS-20B（专家混合模型，总计20.9B参数，约3.61B活跃）进行了单张H100 GPU（bf16）评估，对比密集型基线Qwen3-32B和Yi-34B，在多个维度进行分析。测量了首次生成时间（TTFT）、完整解码吞吐量（TPOT）、端到端延迟分位数、持有过去键值（PKV）时的峰值显存及能量消耗。在2048令牌上下文、64令牌解码条件下，GPT-OSS-20B解码吞吐量与每焦耳生成令牌数均优于两个密集基线，同时显著降低峰值显存和每1000个生成令牌的能耗；其首次生成时间较高，源于MoE路由开销。仅激活17.3%参数（3.61B/20.9B），GPT-OSS-20B相比Qwen3-32B在相同条件下解码吞吐量高出31.8%，每1000令牌能耗低25.8%，峰值显存少31.7%。按活跃参数归一化后，其每活跃参数效率（APE）明显更优，凸显了MoE在部署上的优势。本研究不涉及准确率评估，聚焦部署性能。代码与结果已发布，支持复现与扩展。
原文摘要 · Abstract (English)
We present a single-GPU (H100, bf16) evaluation of GPT-OSS-20B (Mixture-of-Experts; 20.9B total, approx. 3.61B active) against dense baselines Qwen3-32B and Yi-34B across multiple dimensions. We measure true time-to-first-token (TTFT), full-decode throughput (TPOT), end-to-end latency percentiles, peak VRAM with past key values (PKV) held, and energy via a consistent nvidia-smi-based sampler. At a 2048-token context with 64-token decode, GPT-OSS-20B delivers higher decode throughput and tokens per Joule than dense baselines Qwen3-32B and Yi-34B, while substantially reducing peak VRAM and energy per 1000 generated tokens; its TTFT is higher due to MoE routing overhead. With only 17.3% of parameters active (3.61B of 20.9B), GPT-OSS-20B provides about 31.8% higher decode throughput and 25.8% lower energy per 1000 generated tokens than Qwen3-32B at 2048/64, while using 31.7% less peak VRAM. Normalized by active parameters, GPT-OSS-20B shows markedly stronger per-active-parameter efficiency (APE), underscoring MoE's deployment advantages. We do not evaluate accuracy; this is a deployment-focused study. We release code and consolidated results to enable replication and extension.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。