diff --git a/src/main/scala/rocketchip/Periphery.scala b/src/main/scala/rocketchip/Periphery.scala
index 92bbbd76..2fac8023 100644
--- a/src/main/scala/rocketchip/Periphery.scala
+++ b/src/main/scala/rocketchip/Periphery.scala
@@ -134,12 +134,16 @@ trait PeripheryMasterAXI4Mem {
       beatBytes = config.beatBytes)
   })
 
-  private val converter = LazyModule(new TLToAXI4(config.idBits))
+  private val converter = LazyModule(new TLToAXI4(config.beatBytes))
+  private val trim = LazyModule(new AXI4IdIndexer(config.idBits))
+  private val yank = LazyModule(new AXI4UserYanker)
   private val buffer = LazyModule(new AXI4Buffer)
 
   mem foreach { case xbar =>
     converter.node := xbar.node
-    buffer.node := converter.node
+    trim.node := converter.node
+    yank.node := trim.node
+    buffer.node := yank.node
     mem_axi4 := buffer.node
   }
 }
@@ -207,10 +211,12 @@ trait PeripheryMasterAXI4MMIO {
 
   mmio_axi4 :=
     AXI4Buffer()(
+    AXI4UserYanker()(
     AXI4Deinterleaver(cacheBlockBytes)(
-    TLToAXI4(idBits = config.idBits)(      // use idBits = 0 for AXI4-Lite
+    AXI4IdIndexer(config.idBits)(
+    TLToAXI4(config.beatBytes)(
     TLWidthWidget(socBusConfig.beatBytes)( // convert width before attaching to socBus
-    socBus.node))))
+    socBus.node))))))
 }
 
 trait PeripheryMasterAXI4MMIOBundle {
@@ -241,7 +247,7 @@ trait PeripherySlaveAXI4 extends HasTopLevelNetworks {
   fsb.node :=
     TLWidthWidget(config.beatBytes)(
     AXI4ToTL()(
-    AXI4UserYanker(1 << (config.sourceBits - fifoBits - 1))(
+    AXI4UserYanker(Some(1 << (config.sourceBits - fifoBits - 1)))(
     AXI4Fragmenter()(
     AXI4IdIndexer(fifoBits)(
     l2FrontendAXI4Node)))))
diff --git a/src/main/scala/uncore/axi4/Test.scala b/src/main/scala/uncore/axi4/Test.scala
index 0d2a3777..2e75d370 100644
--- a/src/main/scala/uncore/axi4/Test.scala
+++ b/src/main/scala/uncore/axi4/Test.scala
@@ -26,8 +26,8 @@ class AXI4LiteFuzzRAM()(implicit p: Parameters) extends LazyModule
 
   model.node := fuzz.node
   xbar.node  := TLDelayer(0.1)(TLBuffer(BufferParams.flow)(TLDelayer(0.2)(model.node)))
-  ram.node   := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(0, true )(xbar.node)))
-  gpio.node  := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(0, false)(xbar.node)))
+  ram.node   := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4, true )(xbar.node)))
+  gpio.node  := AXI4Fragmenter()(AXI4Deinterleaver(16)(TLToAXI4(4, false)(xbar.node)))
 
   lazy val module = new LazyModuleImp(this) with HasUnitTestIO {
     io.finished := fuzz.module.io.finished
@@ -104,7 +104,7 @@ class AXI4FuzzSlave()(implicit p: Parameters) extends LazyModule
     TLBuffer(BufferParams.flow)(
     TLDelayer(0.1)(
     AXI4ToTL()(
-    AXI4UserYanker(4)(
+    AXI4UserYanker(Some(4))(
     AXI4Fragmenter()(
     AXI4IdIndexer(2)(
     node))))))))
diff --git a/src/main/scala/uncore/axi4/UserYanker.scala b/src/main/scala/uncore/axi4/UserYanker.scala
index 286e4a24..521b04a2 100644
--- a/src/main/scala/uncore/axi4/UserYanker.scala
+++ b/src/main/scala/uncore/axi4/UserYanker.scala
@@ -8,8 +8,10 @@ import config._
 import diplomacy._
 import uncore.tilelink2.UIntToOH1
 
-class AXI4UserYanker(maxFlightPerId: Int)(implicit p: Parameters) extends LazyModule
+class AXI4UserYanker(capMaxFlight: Option[Int] = None)(implicit p: Parameters) extends LazyModule
 {
+  // !!! make maxFlightPerId a cap and maxFlight a per AXI4 Master parameter
+  val maxFlightPerId = capMaxFlight.getOrElse(8)
   require (maxFlightPerId >= 1)
 
   val node = AXI4AdapterNode(
@@ -80,8 +82,8 @@ class AXI4UserYanker(maxFlightPerId: Int)(implicit p: Parameters) extends LazyMo
 object AXI4UserYanker
 {
   // applied to the AXI4 source node; y.node := AXI4UserYanker(idBits, maxFlight)(x.node)
-  def apply(maxFlight: Int)(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
-    val yanker = LazyModule(new AXI4UserYanker(maxFlight))
+  def apply(capMaxFlight: Option[Int] = None)(x: AXI4OutwardNode)(implicit p: Parameters, sourceInfo: SourceInfo): AXI4OutwardNode = {
+    val yanker = LazyModule(new AXI4UserYanker(capMaxFlight))
     yanker.node := x
     yanker.node
   }