Unverified50% confidenceFactExact time
SigLIP2 SO400M采用ViT-SO400M架构，参数量约4亿，输出嵌入维度为1152
1
Sources
50%
Confidence
Long-term
Relevance
7/10/2026
First Seen
Sources
DINOv2 vs SigLIP：k-NN分类为何差距悬殊？视觉编码器选型避坑指南
redditr/MachineLearning7/8/2026
Related Claims
UnverifiedDINOv2 Giant参数量约11亿、嵌入维度1536，均显著大于SigLIP2 SO400M（约4亿参数、维度1152）和CLIP ViT-L（约3亿参数、维度768），但k-NN任务表现最差69% similarUnverified消费级 GPU 如 RTX 4090 拥有 24 GB 显存，可运行数十亿参数的量化模型67% similarPartially VerifiedA100 80GB采用HBM2e显存，带宽约2039GB/s；H100 SXM采用HBM3，带宽约3.35TB/s，FP64 Tensor Core算力约67 TFLOPS，是A100（约19.5 TFLOPS）的3倍以上65% similarVerifiedESP32搭载双核Xtensa LX6处理器，主频可达240MHz，拥有520KB SRAM64% similarUnverified英伟达H100 GPU单卡功耗高达700瓦，一个配备5万块H100的超算集群总功耗可达35兆瓦（MW）64% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/469426API
curl https://kongchang.com/api/v1/knowledge/claims/469426MCP
get_claim(id=469426)