diff --git a/samples/cxl_mem_tg/cxl_mem_tg.h b/samples/cxl_mem_tg/cxl_mem_tg.h
index 03b335d8bdc6..94a6f22f396c 100644
--- a/samples/cxl_mem_tg/cxl_mem_tg.h
+++ b/samples/cxl_mem_tg/cxl_mem_tg.h
@@ -41,7 +41,7 @@ static const uint64_t KB = 1024;
 static const uint64_t MB = KB * 1024;
 static const uint64_t GB = MB * 1024;
 static const uint64_t FPGA_32GB_CACHE_LINES = (32 * GB) / 64;
-static const uint64_t MEM_TG_TEST_TIMEOUT = 3000000;
+static const uint64_t MEM_TG_TEST_TIMEOUT = 500000;
 static const uint64_t TEST_SLEEP_INVL = 100;
 static const uint64_t TG_CTRL_CLEAR = 0x8000000000000000;
 static const uint64_t TG_SLEEP = 300 / 1000;
@@ -292,5 +292,19 @@ class cxl_mem_tg : public test_afu {
   }
 
   token::ptr_t get_token() { return handle_->get_token(); }
+
+  bool option_passed(string str) {
+      CLI::Option* opt = app_.get_option_no_throw(str);
+      if (opt && opt->count() == 1) {
+          return true;
+      }
+      return false;
+  }
+
+  uint64_t get_timeout() {
+      return timeout_msec_;
+  }
+
+
 };
 }  // end of namespace cxl_mem_tg
diff --git a/samples/cxl_mem_tg/cxl_tg_test.h b/samples/cxl_mem_tg/cxl_tg_test.h
index a8b0a557b554..55a35b17c4ed 100644
--- a/samples/cxl_mem_tg/cxl_tg_test.h
+++ b/samples/cxl_mem_tg/cxl_tg_test.h
@@ -34,6 +34,7 @@
 #include <string>
 #include <thread>
 #include <vector>
+#include <ofs/ofs_primitives.h>
 
 #include "afu_test.h"
 #include "cxl_mem_tg.h"
@@ -74,7 +75,7 @@ namespace cxl_mem_tg {
 
 class cxl_tg_test : public test_command {
  public:
-  cxl_tg_test() : tg_offset_(0x0), tg_exe_(NULL) {}
+  cxl_tg_test() : tg_offset_(0x0), timeout_usec_(MEM_TG_TEST_TIMEOUT), tg_exe_(NULL) {}
 
   virtual ~cxl_tg_test() {}
 
@@ -92,6 +93,34 @@ class cxl_tg_test : public test_command {
            ((1000.0 / (double)(tg_exe_->mem_speed_/1000) * (double)num_ticks));
   }
 
+  int ofs_wait_for_eq32(uint32_t offset, uint32_t value,
+      uint64_t timeout_usec, uint32_t sleep_usec) {
+      OFS_TIMESPEC_USEC(ts, sleep_usec);
+      struct timespec begin, now, save, rem;
+      save = ts;
+      uint32_t csr;
+      csr = tg_exe_->read32(offset);
+      clock_gettime(CLOCK_MONOTONIC, &begin);
+      while (csr != value) {
+          if (sleep_usec) {
+              ts = save;
+              while ((nanosleep(&ts, &rem) == -1) &&
+                  (errno == EINTR))
+                  ts = rem;
+          }
+          csr = tg_exe_->read32(offset);
+          clock_gettime(CLOCK_MONOTONIC, &now);
+          struct timespec delta;
+          ofs_diff_timespec(&delta, &now, &begin);
+          uint64_t delta_nsec = delta.tv_nsec + delta.tv_sec * SEC2NSEC;
+          if (delta_nsec > timeout_usec * USEC2NSEC) {
+              return 1;
+          }
+      }
+      return 0;
+
+  }
+
   void print_he_mem_tg() {
     tg_exe_->logger_->debug("DFH:0x{:x}", tg_exe_->read64(AFU_DFH));
     tg_exe_->logger_->debug("GUIDL:0x{:x}", tg_exe_->read64(AFU_ID_L));
@@ -215,20 +244,16 @@ class cxl_tg_test : public test_command {
     uint32_t tg_fail = 0;
     tg_exe_->logger_->debug("tg wait for test completion...");
 
-    tg_status = tg_exe_->read32(TG_TEST_COMPLETE);
-    tg_exe_->logger_->debug("test complete status:{}", tg_status);
-
-    while (tg_status != 0x1) {
-      tg_status = tg_exe_->read32(TG_TEST_COMPLETE);
-      std::this_thread::yield();
-      if (--timeout == 0) {
+    int ret = ofs_wait_for_eq32(TG_TEST_COMPLETE, 0x1,
+        timeout_usec_, TEST_SLEEP_INVL);
+    if (ret != 0) {
         std::cerr << "test completion timeout " << std::endl;
         tg_exe_->status_ = -1;
         tg_print_fail_info();
         return false;
-      }
     }
 
+    tg_exe_->logger_->debug("test complete status:{}", tg_exe_->read32(TG_TEST_COMPLETE));
     tg_exe_->logger_->debug("tg pass:{}", tg_exe_->read32(TG_PASS));
 
     tg_fail = tg_exe_->read32(TG_FAIL);
@@ -303,7 +328,13 @@ class cxl_tg_test : public test_command {
         return -1;
     }
 
-    if (tg_exe_->rcnt_ > tg_exe_->wcnt_) {
+    if (tg_exe_->option_passed("--timeout")) {
+        timeout_usec_ = tg_exe_->get_timeout() * 1000;
+    }
+
+    if (tg_exe_->option_passed("--writes") &&
+        tg_exe_->option_passed("--reads") &&
+        (tg_exe_->rcnt_ > tg_exe_->wcnt_)) {
         cerr << "Read count exceeds Write count" << endl;
         return -1;
     }
@@ -448,6 +479,7 @@ class cxl_tg_test : public test_command {
 
  protected:
   uint64_t tg_offset_;
+  uint64_t timeout_usec_;
   cxl_mem_tg *tg_exe_;
 };