Unverified50% confidenceBenchmarkExact time
Nsight Systems profiling shows that at 7.3 tok/s each token takes about 137ms, of which the GPU is busy about 92ms and idle about 54ms, with the inter-launch gap only about 13ms, so CUDA graphs can only optimize those 13ms.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
10/17/2026
First Seen
Valid until: 1/15/2027
Sources
Related Entities
Related Claims
UnverifiedIn the test, model loading took only a few seconds, GPU usage dropped from 23GB to about 17GB, the first test prompt processing took 225ms, generation speed was stable at 33.5 tok/s, with a total output of 1500 tokens.74% similarVerifiedTaking a 70B parameter model with FP16 precision and 32 attention head layers as an example, the KV Cache for a single 200K-length request requires approximately 80–120 GB of VRAM, far exceeding the 80 GB VRAM limit of a single A100 GPU.73% similarUnverifiedNVIDIA CUDA运行时初始化时创建的设备上下文通常占用150-300MB显存69% similarUnverifiedRunning a 16-bit precision 70B model on 2×80GB GPUs leaves only 2.9 GB of VRAM after the weights are loaded, accommodating only about 9,000 in-flight tokens, which in practice cannot serve requests effectively.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1053040API
curl https://kongchang.com/api/v1/knowledge/claims/1053040MCP
get_claim(id=1053040)