diff --git a/arb/ui_pow_ui.c b/arb/ui_pow_ui.c
index 826cc196..48e69346 100644
--- a/arb/ui_pow_ui.c
+++ b/arb/ui_pow_ui.c
@@ -292,7 +292,10 @@ arb_ui_pow_ui(arb_t res, ulong a, ulong exp, slong prec)
         }
         else   /* or exactly */
         {
-            mpn_sqr(tmp, yman, yn);
+            if (ARF_USE_FFT_MUL(yn))
+                flint_mpn_mul_fft_main(tmp, yman, yn, yman, yn);
+            else
+                mpn_sqr(tmp, yman, yn);
             yn = 2 * yn;
         }
 
diff --git a/arf.h b/arf.h
index 2c193f52..87da5762 100644
--- a/arf.h
+++ b/arf.h
@@ -826,6 +826,11 @@ void arf_urandom(arf_t x, flint_rand_t state, slong bits, arf_rnd_t rnd);
         r3 += r2 < t1;                                      \
     } while (0)
 
+#define ARF_USE_FFT_MUL(_xn) ((_xn) > 10000)
+
+/* from flint/fft.h */
+void flint_mpn_mul_fft_main(mp_ptr r1, mp_srcptr i1, mp_size_t n1, mp_srcptr i2, mp_size_t n2);
+
 #define ARF_MPN_MUL(_z, _x, _xn, _y, _yn) \
     if ((_xn) == (_yn)) \
     { \
@@ -842,19 +847,19 @@ void arf_urandom(arf_t x, flint_rand_t state, slong bits, arf_rnd_t rnd);
             __arb_y1 = (_y)[1]; \
             nn_mul_2x2((_z)[3], (_z)[2], (_z)[1], (_z)[0], __arb_x1, __arb_x0, __arb_y1, __arb_y0); \
         } \
+        else if (ARF_USE_FFT_MUL(_xn)) \
+            flint_mpn_mul_fft_main((_z), (_x), (_xn), (_y), (_yn)); \
         else if ((_x) == (_y)) \
-        { \
             mpn_sqr((_z), (_x), (_xn)); \
-        } \
         else \
-        { \
             mpn_mul_n((_z), (_x), (_y), (_xn)); \
-        } \
     } \
     else if ((_xn) > (_yn)) \
     { \
         if ((_yn) == 1) \
             (_z)[(_xn) + (_yn) - 1] = mpn_mul_1((_z), (_x), (_xn), (_y)[0]); \
+        else if (ARF_USE_FFT_MUL(_yn)) \
+            flint_mpn_mul_fft_main((_z), (_x), (_xn), (_y), (_yn)); \
         else \
             mpn_mul((_z), (_x), (_xn), (_y), (_yn)); \
     } \
@@ -862,6 +867,8 @@ void arf_urandom(arf_t x, flint_rand_t state, slong bits, arf_rnd_t rnd);
     { \
         if ((_xn) == 1) \
             (_z)[(_xn) + (_yn) - 1] = mpn_mul_1((_z), (_y), (_yn), (_x)[0]); \
+        else if (ARF_USE_FFT_MUL(_xn)) \
+            flint_mpn_mul_fft_main((_z), (_y), (_yn), (_x), (_xn)); \
         else \
             mpn_mul((_z), (_y), (_yn), (_x), (_xn)); \
     }
diff --git a/arf/mul_rnd_any.c b/arf/mul_rnd_any.c
index bdb01e2e..5dbffa41 100644
--- a/arf/mul_rnd_any.c
+++ b/arf/mul_rnd_any.c
@@ -55,6 +55,10 @@ arf_mul_rnd_any(arf_ptr z, arf_srcptr x, arf_srcptr y,
         {
             tmp[zn - 1] = mpn_mul_1(tmp, xptr, xn, yptr[0]);
         }
+        else if (ARF_USE_FFT_MUL(yn))
+        {
+            flint_mpn_mul_fft_main(tmp, xptr, xn, yptr, yn);
+        }
         else if (xn == yn)
         {
             if (xptr == yptr)
diff --git a/arf/mul_rnd_down.c b/arf/mul_rnd_down.c
index 731a1d54..3765b021 100644
--- a/arf/mul_rnd_down.c
+++ b/arf/mul_rnd_down.c
@@ -194,7 +194,11 @@ arf_mul_rnd_down(arf_ptr z, arf_srcptr x, arf_srcptr y, slong prec)
         alloc = zn = xn + yn;
         ARF_MUL_TMP_ALLOC(tmp, alloc)
 
-        if (xn == yn)
+        if (ARF_USE_FFT_MUL(yn))
+        {
+            flint_mpn_mul_fft_main(tmp, xptr, xn, yptr, yn);
+        }
+        else if (xn == yn)
         {
             if (xptr == yptr)
                 mpn_sqr(tmp, xptr, xn);