diff --git a/rdrand.h b/rdrand.h
index 2729ed7f..d5798643 100644
--- a/rdrand.h
+++ b/rdrand.h
@@ -42,10 +42,13 @@ public:
 	//! \param retries the number of retries for failed calls to the hardware
 	//! \details RDRAND() constructs a generator with a maximum number of retires
 	//!   for failed generation attempts.
-	//! \details Empirical testing under a 6th generaton i7 (6200U) shows RDSEED fails
-	//!   to fulfill requests at about 6 to 8 times the rate of RDRAND. The default
-	//!   retries reflects the difference.
-	RDRAND(unsigned int retries = 12) : m_retries(retries) {}
+	//! \details According to DJ of Intel, the Intel RDRAND circuit does not underflow.
+	//!   If it did hypothetically underflow, then it would return 0 for the random value.
+	//!   Its not clear what AMD's behavior will be, and what the returned value will be if
+	//!   underflow occurs.
+	//!   Also see <A HREF="https://lists.randombit.net/pipermail/cryptography/2016-June/007702.html">RDRAND
+	//!   not really random with Oracle Studio 12.3 + patches</A>
+	RDRAND(unsigned int retries = 4) : m_retries(retries) {}
 
 	virtual ~RDRAND() {}