diff --git a/src/47-cuda-events/README.md b/src/47-cuda-events/README.md
index 1927b93..7a48add 100644
--- a/src/47-cuda-events/README.md
+++ b/src/47-cuda-events/README.md
@@ -22,6 +22,8 @@ This is where eBPF comes to the rescue! By using uprobes, we can intercept CUDA
 - Error codes and failures
 - Timing of operations
 
+This blog is mainly focus on the CPU side of the CUDA API calls, for fined-grained tracing of GPU operations, you can see [eGPU](https://dl.acm.org/doi/10.1145/3723851.3726984) paper and [bpftime](https://github.com/eunomia-bpf/bpftime) project.
+
 ## Key CUDA Functions We Trace
 
 Our tracer monitors several critical CUDA functions that represent the main operations in GPU computing. Understanding these functions helps you interpret the tracing results and diagnose issues in your CUDA applications:
diff --git a/src/47-cuda-events/README.zh.md b/src/47-cuda-events/README.zh.md
index c7b76a7..e88f283 100644
--- a/src/47-cuda-events/README.zh.md
+++ b/src/47-cuda-events/README.zh.md
@@ -22,6 +22,8 @@ CUDA（Compute Unified Device Architecture，计算统一设备架构）是NVIDI
 - 错误代码和失败原因
 - 操作的时间信息
 
+本教程主要关注CPU侧的CUDA API调用，对于细粒度的GPU操作追踪，你可以参考[eGPU](https://dl.acm.org/doi/10.1145/3723851.3726984)论文和[bpftime](https://github.com/eunomia-bpf/bpftime)项目。
+
 ## eBPF技术背景与GPU追踪的挑战
 
 eBPF（Extended Berkeley Packet Filter）最初是为网络数据包过滤而设计的，但现在已经发展成为一个强大的内核编程框架，使开发人员能够在内核空间运行用户定义的程序，而无需修改内核源代码或加载内核模块。eBPF的安全性通过静态分析和运行时验证器来保证，这使得它能够在生产环境中安全地运行。