diff --git a/libkineto/src/CudaDeviceProperties.cpp b/libkineto/src/CudaDeviceProperties.cpp
index eda37faf0..d5e9252ab 100644
--- a/libkineto/src/CudaDeviceProperties.cpp
+++ b/libkineto/src/CudaDeviceProperties.cpp
@@ -7,10 +7,7 @@
 
 #include "CudaDeviceProperties.h"
 
-#include <vector>
-
 #include <cuda_runtime.h>
-#include <cuda_occupancy.h>
 
 #include "Logger.h"
 
diff --git a/libkineto/src/CudaDeviceProperties.h b/libkineto/src/CudaDeviceProperties.h
index 8156c8b7d..02886b25c 100644
--- a/libkineto/src/CudaDeviceProperties.h
+++ b/libkineto/src/CudaDeviceProperties.h
@@ -7,10 +7,14 @@
 
 #pragma once
 
+#include <vector>
 #include <stdint.h>
+#include <cuda_occupancy.h>
 
 namespace KINETO_NAMESPACE {
 
+const std::vector<cudaOccDeviceProp>& occDeviceProps();
+
 float kernelOccupancy(
     uint32_t deviceId,
     uint16_t registersPerThread,
diff --git a/libkineto/src/output_json.cpp b/libkineto/src/output_json.cpp
index 50577329e..7e6bb391b 100644
--- a/libkineto/src/output_json.cpp
+++ b/libkineto/src/output_json.cpp
@@ -38,6 +38,37 @@ void ChromeTraceLogger::handleTraceStart(
   "schemaVersion": {},
   )JSON", kSchemaVersion);
 
+#ifdef HAS_CUPTI
+  const std::vector<cudaOccDeviceProp>& occProps = KINETO_NAMESPACE::occDeviceProps();
+  if (occProps.size() > 0) {
+    std::ostringstream oss;
+    oss << "[";
+    for (size_t i = 0; i < occProps.size(); i += 1) {
+      const cudaOccDeviceProp& occProp = occProps[i];
+      if (i > 0) {
+        oss << ", ";
+      }
+      oss << "{";
+      oss << "\"computeMajor\": " << occProp.computeMajor << ", ";
+      oss << "\"computeMinor\": " << occProp.computeMinor << ", ";
+      oss << "\"maxThreadsPerBlock\": " << occProp.maxThreadsPerBlock << ", ";
+      oss << "\"maxThreadsPerMultiprocessor\": " << occProp.maxThreadsPerMultiprocessor << ", ";
+      oss << "\"regsPerBlock\": " << occProp.regsPerBlock << ", ";
+      oss << "\"regsPerMultiprocessor\": " << occProp.regsPerMultiprocessor << ", ";
+      oss << "\"warpSize\": " << occProp.warpSize << ", ";
+      oss << "\"sharedMemPerBlock\": " << occProp.sharedMemPerBlock << ", ";
+      oss << "\"sharedMemPerMultiprocessor\": " << occProp.sharedMemPerMultiprocessor << ", ";
+      oss << "\"numSms\": " << occProp.numSms << ", ";
+      oss << "\"sharedMemPerBlockOptin\": " << occProp.sharedMemPerBlockOptin;
+      oss << "}";
+    }
+    oss << "]";
+    traceOf_ << fmt::format(R"JSON(
+  "cudaOccDeviceProps": {},
+  )JSON", oss.str());
+  }
+#endif // HAS_CUPTI
+
   if (!metadata.empty()) {
     traceOf_ << R"JSON(
   "metadata": {