Unverified50% confidenceBenchmarkExact time
RTX 3060的INT4 Tensor Core理论峰值算力约为101 TOPS，而FP16仅约12.7 TFLOPS，差距达8倍以上
1
Sources
50%
Confidence
Long-term
Relevance
7/12/2026
First Seen
Sources
ComfyUI-INT4-Fast：6GB显卡流畅运行Flux大模型实测
redditr/StableDiffusion7/10/2026
Related Claims
UnverifiedINT4量化将原本由65536个不同值（FP16）表示的权重压缩到仅16个离散级别61% similarUnverified在覆盖九大维度共403道题的测试中，Muse-Glimmer-30B综合均分为76.04，比Corpus 3.6 27B的75.57高出0.47分61% similarUnverifiedMuse Glimmer 30B在RTX 5090上报告可达每秒74.9个token，评测者在RTX 3090上运行全精度版本观察到约60-65 token/秒60% similarUnverifiedBlackwell架构FP4矩阵乘法理论峰值算力相比FP16提升可达8倍60% similarUnverifiedQwen3.6 27B NVFP4在工具调用（ToolCore）测试中获得满分100分60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/489323API
curl https://kongchang.com/api/v1/knowledge/claims/489323MCP
get_claim(id=489323)