diff --git a/cuburn/code/util.py b/cuburn/code/util.py
index a12bcea..b27dc06 100644
--- a/cuburn/code/util.py
+++ b/cuburn/code/util.py
@@ -340,5 +340,5 @@ __device__ uint32_t rb_incr(uint32_t &rb_base, int tidx) {
 # done when I get my hands on a Kepler device. The fixed size assumes blocks
 # of 256 threads, although even at that size there are pathological cases that
 # could break the assumption.
-DEFAULT_RB_SIZE = 256
+DEFAULT_RB_SIZE = 1024
 ringbuflib = mkringbuflib(DEFAULT_RB_SIZE)