diff --git a/ChangeLog b/ChangeLog
index f7a4c9e9..4869fdfd 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -3,6 +3,8 @@
 - oops, ==0 missing from a strcmp() in vips7compat
 - fixed a race in im_XYZ2Lab() table build
 - added im_concurrency_get() to operation db
+- better benchmarkn.sh runs for the correct number of CPUs automatically, runs
+  three times for each one, and just reports the fastest
 
 26/7/11 started 7.26.0
 - version bunp for 7.26
diff --git a/benchmark/benchmarkn.sh b/benchmark/benchmarkn.sh
index 16a40b63..9527cda4 100755
--- a/benchmark/benchmarkn.sh
+++ b/benchmark/benchmarkn.sh
@@ -1,4 +1,4 @@
-#!/bin/sh
+#!/bin/bash
 
 uname -a
 gcc --version
@@ -7,7 +7,7 @@ vips --version
 # how large an image do you want to process? 
 # sample2.v is 290x442 pixels ... replicate this many times horizontally and 
 # vertically to get a highres image for the benchmark
-tile=2
+tile=13
 
 # how complex an operation do you want to run?
 # this sets the number of copies of the benchmark we chain together:
@@ -21,26 +21,44 @@ if [ $? != 0 ]; then
   echo "build of test image failed -- out of disc space?"
   exit 1
 fi
-echo -n "test image is" `header -f Xsize temp.v` 
-echo " by" `header -f Ysize temp.v` "pixels"
+echo -n "test image is" `header -f width temp.v` 
+echo " by" `header -f height temp.v` "pixels"
+max_cpus=`vips im_concurrency_get`
 
+echo "max cpus = $max_cpus"
 echo "starting benchmark ..."
-echo "chain=$chain"
-
-for cpus in 1 2 3 4 ; do
-  export IM_CONCURRENCY=$cpus
-
-  echo IM_CONCURRENCY=$IM_CONCURRENCY
-  echo time -p vips im_benchmarkn temp.v temp2.v $chain
-  time -p vips im_benchmarkn temp.v temp2-$cpus.v $chain
+echo /usr/bin/time -f %e vips \
+  --vips-concurrency=xx \
+  --vips-tile-width=64 --vips-tile-height=64 \
+  im_benchmarkn temp.v temp2.v $chain
+echo reported real-time is best of three runs
+echo cpus real-time
 
+for((cpus = 1; cpus <= max_cpus; cpus++)); do
+  t1=`/usr/bin/time -f %e vips \
+	  --vips-concurrency=$cpus \
+	  --vips-tile-width=64 --vips-tile-height=64 \
+	  im_benchmarkn temp.v temp2.v $chain 2>&1`
   if [ $? != 0 ]; then
     echo "benchmark failed -- install problem?"
     exit 1
   fi
+  t2=`/usr/bin/time -f %e vips \
+	  --vips-concurrency=$cpus \
+	  --vips-tile-width=64 --vips-tile-height=64 \
+	  im_benchmarkn temp.v temp2.v $chain 2>&1`
+  t3=`/usr/bin/time -f %e vips \
+	  --vips-concurrency=$cpus \
+	  --vips-tile-width=64 --vips-tile-height=64 \
+	  im_benchmarkn temp.v temp2.v $chain 2>&1`
 
-  # find pixel average ... should be the same for all IM_CONCURRENCY settings
-  # or we have some kind of terrible bug
-  echo vips im_avg temp2-$cpus.v
-  vips im_avg temp2-$cpus.v
+  # echo $t1 $t2 $t3
+
+  if [[ $t2 < $t1 ]]; then
+	  t1=$t2
+  fi
+  if [[ $t3 < $t1 ]]; then
+	  t1=$t3
+  fi
+  echo $cpus $t1
 done