diff --git a/docs/Details_Usage.md b/docs/Details_Usage.md
index 51560acd01b62fc5df98f115e591aeaf859712bd..d2bfa43873749a5fd93eaa866793ee7467c62dc8 100644
--- a/docs/Details_Usage.md
+++ b/docs/Details_Usage.md
@@ -930,8 +930,8 @@ PmuDeviceData *devData = nullptr;
 auto len = PmuGetDevMetric(oriData, oriLen, devAttr, 1, &devData);
 // devData的长度等于cluster个数
 for (int i=0;i<len;++i) {
-    // 每个devData表示一个cluster的L3平均时延，是以cycles为单位
-    std::cout << "L3 latency(" << devData[i].clusterId << "): " << devData[i].count<< " cycles\n";
+    // 每个devData表示一个cluster的L3平均时延，是以ns为单位
+    cout << "L3 latency(" << devData[i].clusterId << "): " << devData[i].count<< " ns\n";
 }
 DevDataFree(devData);
 PmuDataFree(oriData);
@@ -953,7 +953,7 @@ kperf.disable(pd)
 ori_data = kperf.read(pd)
 dev_data = kperf.get_device_metric(ori_data, dev_attr)
 for data in dev_data.iter:
-    print(f"L3 latency({data.clusterId}): {data.count} cycles")
+    print(f"L3 latency({data.clusterId}): {data.count} ns")
 ```
 
 ```go
@@ -970,7 +970,7 @@ kperf.PmuDisable(fd)
 dataVo, _ := kperf.PmuRead(fd)
 deivceDataVo, _ := kperf.PmuGetDevMetric(dataVo, deviceAttrs)
 for _, v := range deivceDataVo.GoDeviceData {
-	fmt.Printf("L3 latency(%v): %v cycles\n", v.ClusterId, v.Count)
+	fmt.Printf("L3 latency(%v): %v ns\n", v.ClusterId, v.Count)
 }
 kperf.DevDataFree(deivceDataVo)
 kperf.PmuDataFree(dataVo)
@@ -979,10 +979,10 @@ kperf.PmuClose(fd)
 
 执行上述代码，输出的结果类似如下：
 ```
-L3 latency(0): 101 cycles
-L3 latency(1): 334.6 cycles
-L3 latency(2): 267.8 cycles
-L3 latency(3): 198.4 cycles
+L3 latency(0): 101 ns
+L3 latency(1): 334.6 ns
+L3 latency(2): 267.8 ns
+L3 latency(3): 198.4 ns
 ...
 ```
 
diff --git a/include/pmu.h b/include/pmu.h
index 0873c0caf5600dda429f436f3c8bcc7bed69d90b..2c7dd3b4df26656058e80cd313dac68820900627 100644
--- a/include/pmu.h
+++ b/include/pmu.h
@@ -442,7 +442,7 @@ enum PmuDeviceMetric {
     PMU_L3_REF,
     // Percluster metric.
     // Collect L3 total latency for each cluster node.
-    // Unit: cycles
+    // Unit: ns 
     PMU_L3_LAT,
     // Collect pcie rx bandwidth.
     // Perpcie metric.
diff --git a/pmu/pmu_metric.cpp b/pmu/pmu_metric.cpp
index be581c14f0a2201c143d2c623211d2c45762dec6..64c9908f88c29e55b3fb00ceaabda4d13d388e85 100644
--- a/pmu/pmu_metric.cpp
+++ b/pmu/pmu_metric.cpp
@@ -324,6 +324,8 @@ namespace KUNPENG_PMU {
         METRIC_CONFIG::HHA_CROSS_SOCKET,
     };
 
+    static const map<CHIP_TYPE, double> L3_CLOCK_NS {{CHIP_TYPE::HIPB, 0.3448275862}};
+
     const UNCORE_METRIC_MAP UNCORE_METRIC_CONFIG_MAP = {
         {CHIP_TYPE::HIPA, HIP_A_UNCORE_METRIC_MAP},
         {CHIP_TYPE::HIPB, HIP_B_UNCORE_METRIC_MAP},
@@ -333,7 +335,7 @@ namespace KUNPENG_PMU {
     {
         CHIP_TYPE chipType = GetCpuType();
         if (UNCORE_METRIC_CONFIG_MAP.find(chipType) == UNCORE_METRIC_CONFIG_MAP.end()) {
-            return {}; 
+            return {};
         }
         return UNCORE_METRIC_CONFIG_MAP.at(chipType);
     }
@@ -958,6 +960,17 @@ namespace KUNPENG_PMU {
         return 64 * rawCount;
     }
 
+    static uint64_t L3Lat(const uint64_t rawCount)
+    {
+        const CHIP_TYPE chipType = GetCpuType();
+        auto iter = L3_CLOCK_NS.find(chipType);
+        uint64_t count = rawCount;
+        if (iter != L3_CLOCK_NS.end()) {
+            count = rawCount * iter->second;
+        }
+        return count;
+    }
+
     static PmuMetricMode GetMetricMode(const PmuDeviceMetric &metric)
     {
         switch(metric) {
@@ -1308,7 +1321,8 @@ namespace KUNPENG_PMU {
 
     map<PmuDeviceMetric, ComputeMetricCb> computeMetricMap = {{PMU_DDR_READ_BW, DDRBw},
                                                                 {PMU_DDR_WRITE_BW, DDRBw},
-                                                                {PMU_L3_TRAFFIC, L3Bw}};
+                                                                {PMU_L3_TRAFFIC, L3Bw},
+                                                                {PMU_L3_LAT, L3Lat}};
     map<PmuDeviceMetric, AggregateMetricCb> aggregateMap = {
         {PMU_DDR_READ_BW, AggregateByChannel},
         {PMU_DDR_WRITE_BW, AggregateByChannel},
diff --git a/third_party/elfin-parser b/third_party/elfin-parser
index fb404f74414a5e07f4317ca865dc5ff04b2c5639..13e57e29400a3a7bb5cb27c364b4b73468b4050f 160000
--- a/third_party/elfin-parser
+++ b/third_party/elfin-parser
@@ -1 +1 @@
-Subproject commit fb404f74414a5e07f4317ca865dc5ff04b2c5639
+Subproject commit 13e57e29400a3a7bb5cb27c364b4b73468b4050f