diff --git a/include/boost/multiprecision/detail/default_ops.hpp b/include/boost/multiprecision/detail/default_ops.hpp
index e278777b..5dd099c4 100644
--- a/include/boost/multiprecision/detail/default_ops.hpp
+++ b/include/boost/multiprecision/detail/default_ops.hpp
@@ -45,6 +45,9 @@ void generic_interconvert(To& to, const From& from, const std::integral_constant
 template <class To, class From>
 void generic_interconvert(To& to, const From& from, const std::integral_constant<int, number_kind_rational>& /*to_type*/, const std::integral_constant<int, number_kind_integer>& /*from_type*/);
 
+template <class Integer>
+BOOST_MP_CXX14_CONSTEXPR Integer karatsuba_sqrt(const Integer& x, Integer& r, Integer& t, size_t bits);
+
 } // namespace detail
 
 namespace default_ops {
@@ -1588,62 +1591,100 @@ inline BOOST_MP_CXX14_CONSTEXPR void eval_bit_unset(T& val, unsigned index)
       eval_bitwise_xor(val, mask);
 }
 
-template <class B>
-void BOOST_MP_CXX14_CONSTEXPR eval_integer_sqrt(B& s, B& r, const B& x)
+template <class Backend>
+BOOST_MP_CXX14_CONSTEXPR void eval_qr(const Backend& x, const Backend& y, Backend& q, Backend& r);
+
+template <class Backend>
+BOOST_MP_CXX14_CONSTEXPR void eval_karatsuba_sqrt(Backend& result, const Backend& x, Backend& r, Backend& t, size_t bits)
 {
-   //
-   // This is slow bit-by-bit integer square root, see for example
-   // http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Binary_numeral_system_.28base_2.29
-   // There are better methods such as http://hal.inria.fr/docs/00/07/28/54/PDF/RR-3805.pdf
-   // and http://hal.inria.fr/docs/00/07/21/13/PDF/RR-4475.pdf which should be implemented
-   // at some point.
-   //
-   using ui_type = typename boost::multiprecision::detail::canonical<unsigned char, B>::type;
+   using default_ops::eval_is_zero;
+   using default_ops::eval_subtract;
+   using default_ops::eval_right_shift;
+   using default_ops::eval_left_shift;
+   using default_ops::eval_bit_set;
+   using default_ops::eval_decrement;
+   using default_ops::eval_bitwise_and;
+   using default_ops::eval_add;
+   using default_ops::eval_qr;
 
-   s = ui_type(0u);
-   if (eval_get_sign(x) == 0)
+   using small_uint = typename std::tuple_element<0, typename Backend::unsigned_types>::type;
+
+   constexpr small_uint zero = 0u;
+
+   // we can calculate it faster with std::sqrt
+#ifdef BOOST_HAS_INT128
+   if (bits <= 128)
    {
-      r = ui_type(0u);
+      unsigned __int128 a, b, c;
+      eval_convert_to(&a, x);
+      c = boost::multiprecision::detail::karatsuba_sqrt(a, b, c, bits);
+      r = number<Backend>::canonical_value(b);
+      result = number<Backend>::canonical_value(c);
       return;
    }
-   int g = eval_msb(x);
-   if (g <= 1)
+#else
+   if (bits <= std::numeric_limits<std::uintmax_t>::digits)
    {
-      s = ui_type(1);
-      eval_subtract(r, x, s);
+      std::uintmax_t a, b, c;
+      eval_convert_to(&a, x);
+      c = boost::multiprecision::detail::karatsuba_sqrt(a, b, c, bits);
+      r = number<Backend>::canonical_value(b);
+      result = number<Backend>::canonical_value(c);
       return;
    }
-
-   B t;
-   r = x;
-   g /= 2;
-   int org_g = g;
-   eval_bit_set(s, g);
-   eval_bit_set(t, 2 * g);
-   eval_subtract(r, x, t);
-   --g;
-   if (eval_get_sign(r) == 0)
-      return;
-   int msbr = eval_msb(r);
-   do
+#endif
+   // https://hal.inria.fr/file/index/docid/72854/filename/RR-3805.pdf
+   std::size_t  b = bits / 4;
+   Backend q(x);
+   eval_right_shift(q, b * 2);
+   Backend s;
+   eval_karatsuba_sqrt(s, q, r, t, bits - b * 2);
+   t = zero;
+   eval_bit_set(t, b * 2);
+   eval_left_shift(r, b);
+   eval_decrement(t);
+   eval_bitwise_and(t, x);
+   eval_right_shift(t, b);
+   eval_add(t, r);
+   eval_left_shift(s, 1u);
+   eval_qr(t, s, q, r);
+   eval_left_shift(r, b);
+   t = zero;
+   eval_bit_set(t, b);
+   eval_decrement(t);
+   eval_bitwise_and(t, x);
+   eval_add(r, t);
+   eval_left_shift(s, b - 1);
+   eval_add(s, q);
+   eval_multiply(q, q);
+   // we substract after, so it works for unsigned integers too
+   if (r.compare(q) < 0)
    {
-      if (msbr >= org_g + g + 1)
-      {
-         t = s;
-         eval_left_shift(t, g + 1);
-         eval_bit_set(t, 2 * g);
-         if (t.compare(r) <= 0)
-         {
-            BOOST_ASSERT(g >= 0);
-            eval_bit_set(s, g);
-            eval_subtract(r, t);
-            if (eval_get_sign(r) == 0)
-               return;
-            msbr = eval_msb(r);
-         }
-      }
-      --g;
-   } while (g >= 0);
+      t = s;
+      eval_left_shift(t, 1u);
+      eval_decrement(t);
+      eval_add(r, t);
+      eval_decrement(s);
+   }
+   eval_subtract(r, q);
+   result = s;
+}
+
+template <class Backend>
+BOOST_MP_CXX14_CONSTEXPR void eval_integer_sqrt(Backend& result, Backend& r, const Backend& x)
+{
+   using small_uint = typename std::tuple_element<0, typename Backend::unsigned_types>::type;
+
+   constexpr small_uint zero = 0u;
+
+   if (eval_is_zero(x))
+   {
+      r = zero;
+      result = zero;
+      return;
+   }
+   Backend t;
+   eval_karatsuba_sqrt(result, x, r, t, eval_msb(x) + 1);
 }
 
 template <class B>
diff --git a/include/boost/multiprecision/integer.hpp b/include/boost/multiprecision/integer.hpp
index eeb6a82b..1db6f613 100644
--- a/include/boost/multiprecision/integer.hpp
+++ b/include/boost/multiprecision/integer.hpp
@@ -189,52 +189,63 @@ BOOST_MP_CXX14_CONSTEXPR typename std::enable_if<boost::multiprecision::detail::
    return val;
 }
 
+namespace detail {
+
 template <class Integer>
-BOOST_MP_CXX14_CONSTEXPR typename std::enable_if<boost::multiprecision::detail::is_integral<Integer>::value, Integer>::type karatsuba_sqrt(const Integer& x, Integer& r, Integer& t, size_t bits)
+BOOST_MP_CXX14_CONSTEXPR Integer karatsuba_sqrt(const Integer& x, Integer& r, Integer& t, size_t bits)
 {
 #ifndef BOOST_MP_NO_CONSTEXPR_DETECTION
-   // std::sqrt is not constexpr by standard, so use this 
-   if (BOOST_MP_IS_CONST_EVALUATED(bits)) {
-      if (bits <= 4) {
-         if (x == 0) {
+   // std::sqrt is not constexpr by standard, so use this
+   if (BOOST_MP_IS_CONST_EVALUATED(bits))
+   {
+      if (bits <= 4)
+      {
+         if (x == 0)
+         {
             r = 0u;
-            return 0u; 
+            return 0u;
          }
-         else if (x < 4) {
+         else if (x < 4)
+         {
             r = x - 1;
-            return 1u; 
+            return 1u;
          }
-         else if (x < 9) {
+         else if (x < 9)
+         {
             r = x - 4;
-            return 2u; 
+            return 2u;
          }
-         else {
+         else
+         {
             r = x - 9;
-            return 3u; 
+            return 3u;
          }
       }
    }
    else
 #endif
    // we can calculate it faster with std::sqrt
-   if (bits <= 64) {
-      const uint64_t int32max = uint64_t((std::numeric_limits<uint32_t>::max)());
-      uint64_t val = static_cast<uint64_t>(x);
-      uint64_t s64 = static_cast<uint64_t>(std::sqrt(static_cast<long double>(val)));
+   if (bits <= 64)
+   {
+      const std::uint64_t int32max = std::uint64_t((std::numeric_limits<std::uint32_t>::max)());
+      std::uint64_t       val      = static_cast<std::uint64_t>(x);
+      std::uint64_t       s64      = static_cast<std::uint64_t>(std::sqrt(static_cast<long double>(val)));
       // converting to long double can loose some precision, and `sqrt` can give eps error, so we'll fix this
       // this is needed
-      while (s64 > int32max || s64 * s64 > val) s64--;
+      while (s64 > int32max || s64 * s64 > val)
+         s64--;
       // in my tests this never fired, but theoretically this might be needed
-      while (s64 < int32max && (s64 + 1) * (s64 + 1) <= val) s64++;
+      while (s64 < int32max && (s64 + 1) * (s64 + 1) <= val)
+         s64++;
       r = val - s64 * s64;
       return s64;
    }
    // https://hal.inria.fr/file/index/docid/72854/filename/RR-3805.pdf
-   size_t b = bits / 4;
+   std::size_t b = bits / 4;
    Integer q = x;
    q >>= b * 2;
    Integer s = karatsuba_sqrt(q, r, t, bits - b * 2);
-   t = 0u;
+   t         = 0u;
    bit_set(t, b * 2);
    r <<= b;
    t--;
@@ -253,7 +264,8 @@ BOOST_MP_CXX14_CONSTEXPR typename std::enable_if<boost::multiprecision::detail::
    s += q;
    q *= q;
    // we substract after, so it works for unsigned integers too
-   if (r < q) {
+   if (r < q)
+   {
       t = s;
       t <<= 1;
       t--;
@@ -264,6 +276,8 @@ BOOST_MP_CXX14_CONSTEXPR typename std::enable_if<boost::multiprecision::detail::
    return s;
 }
 
+} // namespace detail
+
 template <class Integer>
 BOOST_MP_CXX14_CONSTEXPR typename std::enable_if<boost::multiprecision::detail::is_integral<Integer>::value, Integer>::type sqrt(const Integer& x, Integer& r)
 {
@@ -272,7 +286,7 @@ BOOST_MP_CXX14_CONSTEXPR typename std::enable_if<boost::multiprecision::detail::
       return 0u;
    }
    Integer t{};
-   return karatsuba_sqrt(x, r, t, msb(x) + 1);
+   return detail::karatsuba_sqrt(x, r, t, msb(x) + 1);
 }
 
 template <class Integer>
diff --git a/performance/sqrt_bench.cpp b/performance/sqrt_bench.cpp
new file mode 100644
index 00000000..6dd6413c
--- /dev/null
+++ b/performance/sqrt_bench.cpp
@@ -0,0 +1,144 @@
+//  Copyright 2020 John Maddock. Distributed under the Boost
+//  Software License, Version 1.0. (See accompanying file
+//  LICENSE_1_0.txt or copy at https://www.boost.org/LICENSE_1_0.txt
+
+#include <iostream>
+#include <benchmark/benchmark.h>
+#include <boost/multiprecision/cpp_int.hpp>
+#include <boost/multiprecision/gmp.hpp>
+#include <boost/multiprecision/integer.hpp>
+#include <boost/random.hpp>
+#include <cmath>
+
+#include <immintrin.h>
+
+using namespace boost::multiprecision;
+using namespace boost::random;
+
+template <class Integer>
+BOOST_MP_CXX14_CONSTEXPR Integer sqrt_old(const Integer& x, Integer& r)
+{
+   //
+   // This is slow bit-by-bit integer square root, see for example
+   // http://en.wikipedia.org/wiki/Methods_of_computing_square_roots#Binary_numeral_system_.28base_2.29
+   // There are better methods such as http://hal.inria.fr/docs/00/07/28/54/PDF/RR-3805.pdf
+   // and http://hal.inria.fr/docs/00/07/21/13/PDF/RR-4475.pdf which should be implemented
+   // at some point.
+   //
+   Integer s = 0;
+   if (x == 0)
+   {
+      r = 0;
+      return s;
+   }
+   int g = msb(x);
+   if (g == 0)
+   {
+      r = 1;
+      return s;
+   }
+
+   Integer t = 0;
+   r         = x;
+   g /= 2;
+   bit_set(s, g);
+   bit_set(t, 2 * g);
+   r = x - t;
+   --g;
+   do
+   {
+      t = s;
+      t <<= g + 1;
+      bit_set(t, 2 * g);
+      if (t <= r)
+      {
+         bit_set(s, g);
+         r -= t;
+      }
+      --g;
+   } while (g >= 0);
+   return s;
+}
+
+template <class Integer>
+BOOST_MP_CXX14_CONSTEXPR Integer sqrt_old(const Integer& x)
+{
+   Integer r(0);
+   return sqrt_old(x, r);
+}
+
+template <class T>
+std::tuple<std::vector<T>, std::vector<T> >& get_test_vector(unsigned bits)
+{
+   static std::map<unsigned, std::tuple<std::vector<T>, std::vector<T> > > data;
+
+   std::tuple<std::vector<T>, std::vector<T> >& result = data[bits];
+
+   if (std::get<0>(result).size() == 0)
+   {
+      mt19937                     mt;
+      uniform_int_distribution<T> ui(T(1) << (bits - 1), T(1) << bits);
+
+      std::vector<T>& a = std::get<0>(result);
+      std::vector<T>& b = std::get<1>(result);
+
+      for (unsigned i = 0; i < 1000; ++i)
+      {
+         a.push_back(ui(mt));
+         b.push_back(0);
+      }
+   }
+   return result;
+}
+
+template <class T>
+std::vector<T>& get_test_vector_a(unsigned bits)
+{
+   return std::get<0>(get_test_vector<T>(bits));
+}
+template <class T>
+std::vector<T>& get_test_vector_b(unsigned bits)
+{
+   return std::get<1>(get_test_vector<T>(bits));
+}
+
+template <typename T>
+static void BM_sqrt_old(benchmark::State& state)
+{
+   int                         bits = state.range(0);
+
+   std::vector<T>& a = get_test_vector_a<T>(bits);
+   std::vector<T>& b = get_test_vector_b<T>(bits);
+
+   for (auto _ : state)
+   {
+      for (unsigned i = 0; i < a.size(); ++i)
+         b[i] = sqrt_old(a[i]);
+   }
+   state.SetComplexityN(bits);
+}
+
+template <typename T>
+static void BM_sqrt_current(benchmark::State& state)
+{
+   int bits = state.range(0);
+
+   std::vector<T>& a = get_test_vector_a<T>(bits);
+   std::vector<T>& b = get_test_vector_b<T>(bits);
+
+   for (auto _ : state)
+   {
+      for (unsigned i = 0; i < a.size(); ++i)
+         b[i] = sqrt(a[i]);
+   }
+   state.SetComplexityN(bits);
+}
+
+constexpr unsigned lower_range = 512;
+constexpr unsigned upper_range = 1 << 15;
+
+BENCHMARK_TEMPLATE(BM_sqrt_old, cpp_int)->RangeMultiplier(2)->Range(lower_range, upper_range)->Unit(benchmark::kMillisecond)->Complexity();
+BENCHMARK_TEMPLATE(BM_sqrt_current, cpp_int)->RangeMultiplier(2)->Range(lower_range, upper_range)->Unit(benchmark::kMillisecond)->Complexity();
+BENCHMARK_TEMPLATE(BM_sqrt_current, mpz_int)->RangeMultiplier(2)->Range(lower_range, upper_range)->Unit(benchmark::kMillisecond)->Complexity();
+
+BENCHMARK_MAIN();