Unverified60% confidenceFactExact time
Large model inference is typically memory-bandwidth-bound; the computational intensity of llama-2-70B inference is approximately 1–2 FLOP/Byte, far below the H100 peak computational intensity threshold of approximately 120 FLOP/Byte.
2
Sources
60%
Confidence
Long-term
Relevance
7/21/2026
First Seen
Sources
Related Claims
Unverified大语言模型自回归生成阶段属于内存带宽受限任务，单batch推理算术强度约1-2 FLOP/Byte，远低于RTX 4090的计算/带宽比约160 FLOP/Byte78% similarUnverifiedH100的屋顶线约为300 FLOPS/Byte，自回归推理算术强度约1-2 FLOPS/Byte，而DiffusionGemma通过并行处理将算术强度提升约256倍，从memory-bound转为compute-bound73% similarVerifiedStable Diffusion Checkpoint models are typically 2–7GB, while LoRA models are typically only tens to hundreds of MB.66% similarUnverifiedMoE模型采用稀疏激活，Qwen 3.5 122B总参数122B但每次推理仅激活约10B，需将全部参数加载进内存，对内存容量极度敏感66% similarPartially VerifiedQuantization compresses model weights from high-precision floating point numbers such as FP32 to low-precision representations such as INT8 or INT4, reducing memory footprint and computational requirements66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/575666API
curl https://kongchang.com/api/v1/knowledge/claims/575666MCP
get_claim(id=575666)