diff --git a/src/main/scala/radiance/tile/RadianceCluster.scala b/src/main/scala/radiance/tile/RadianceCluster.scala
index 1e39007..e0b7c38 100644
--- a/src/main/scala/radiance/tile/RadianceCluster.scala
+++ b/src/main/scala/radiance/tile/RadianceCluster.scala
@@ -35,7 +35,6 @@ class RadianceCluster (
   // Instantiate cluster-local shared memory scratchpad
   //
   // Instantiate the same number of banks as there are lanes.
-  // val numLsuLanes = 4 // FIXME: hardcoded
 
  // must toSeq here, otherwise Iterable is lazy and will break diplomacy
   val gemminiTiles = leafTiles.values.filter(_.isInstanceOf[GemminiTile]).toSeq.asInstanceOf[Seq[GemminiTile]]
@@ -43,7 +42,7 @@ class RadianceCluster (
 
   // TODO: this probably needs to be instantiated inside the radiance shared mem module
   val virgoSharedMemComponents = new VirgoSharedMemComponents(thisClusterParams, gemminiTiles, radianceTiles)
-  lazy val sharedMemSystem = LazyModule(new RadianceSharedMem(virgoSharedMemComponents, clbus))
+  val sharedMemSystem = LazyModule(new RadianceSharedMem(virgoSharedMemComponents, clbus))
 
   val numCoresInCluster = leafTiles.size - gemminiTiles.size
 
diff --git a/src/main/scala/radiance/tile/RadianceSharedMem.scala b/src/main/scala/radiance/tile/RadianceSharedMem.scala
index 10867bd..76a7e4c 100644
--- a/src/main/scala/radiance/tile/RadianceSharedMem.scala
+++ b/src/main/scala/radiance/tile/RadianceSharedMem.scala
@@ -248,23 +248,14 @@ class RadianceSharedMemImp(outer: RadianceSharedMem) extends LazyModuleImp(outer
     wNode.d <> Queue(writeResp, 2)
   }
 
-
-  // TODO: simplify wire initialization and use tree reduction
-  // read OR write access counter for smem banks
-  val smemBankMgrsGrouped = outer.smemBankMgrs.grouped(outer.smemSubbanks)
-  val numBanks = smemBankMgrsGrouped.length
-  val counterWidth = 32
-  val smemReadsPerBankPerCycle  = Seq.fill(numBanks)(Seq.fill(outer.smemSubbanks)
-  (Wire(UInt(counterWidth.W))))
-  val smemWritesPerBankPerCycle = Seq.fill(numBanks)(Seq.fill(outer.smemSubbanks)
-  (Wire(UInt(counterWidth.W))))
-  val smemReadsPerCycle  = smemReadsPerBankPerCycle.map(_.reduce(_ + _)).reduce(_ + _)
-  val smemWritesPerCycle = smemWritesPerBankPerCycle.map(_.reduce(_ + _)).reduce(_ + _)
-  val smemReadCounter = RegInit(UInt(counterWidth.W), 0.U)
-  val smemWriteCounter = RegInit(UInt(counterWidth.W), 0.U)
-  smemReadCounter  := smemReadCounter + smemReadsPerCycle
-  smemWriteCounter := smemWriteCounter + smemWritesPerCycle
-  // smemReadsPerBankPerCycle.foreach(_.foreach(dontTouch(_)))
+  // read/write access counter for smem banks
+  val Seq(smemReadsPerCycle, smemWritesPerCycle) = outer.smemBankMgrs.transpose.map { rw =>
+    VecInit(rw.map(_.in.head._1.a.fire.asUInt)).reduceTree(_ +& _)
+  }
+  val smemReadCounter = RegInit(0.U(32.W))
+  val smemWriteCounter = RegInit(0.U(32.W))
+  smemReadCounter := smemReadCounter +& smemReadsPerCycle
+  smemWriteCounter := smemWriteCounter +& smemWritesPerCycle
   dontTouch(smemReadCounter)
   dontTouch(smemWriteCounter)
 
@@ -315,11 +306,6 @@ class RadianceSharedMemImp(outer: RadianceSharedMem) extends LazyModuleImp(outer
 
         makeBuffer(mem, rNode, rEdge, wNode, wEdge)
 
-        // TODO: these should also work for non-stride-by-word
-        // add access counters to banks
-        smemReadsPerBankPerCycle(bid)(wid)  := (rNode.a.fire === true.B)
-        smemWritesPerBankPerCycle(bid)(wid) := (wNode.a.fire === true.B)
-
         (uniformFires zip outer.uniformNodesOut).foreach { case (uf, n) =>
           uf(bid)(wid) := n(bid)(wid).in.head._1.a.fire
         }
diff --git a/src/main/scala/radiance/tile/VirgoSharedMemComponents.scala b/src/main/scala/radiance/tile/VirgoSharedMemComponents.scala
index 6c89e94..9e8f8df 100644
--- a/src/main/scala/radiance/tile/VirgoSharedMemComponents.scala
+++ b/src/main/scala/radiance/tile/VirgoSharedMemComponents.scala
@@ -31,7 +31,7 @@ class VirgoSharedMemComponents(
   val gemminis = gemminiTiles.map(_.gemmini)
   val gemminiConfigs = gemminis.map(_.config)
   gemminiConfigs.foreach { config =>
-    assert(smemBanks == config.sp_banks && isPow2(smemBanks / config.sp_banks)) // TODO: should allow >=
+    assert(smemBanks == config.sp_banks && isPow2(smemBanks / config.sp_banks))
     assert(smemWidth >= (config.sp_width / 8) && isPow2(smemWidth / (config.sp_width / 8)))
     assert(smemSize == config.sp_capacity.asInstanceOf[CapacityInKilobytes].kilobytes * 1024)
   }