diff --git a/rocket/src/main/scala/icache.scala b/rocket/src/main/scala/icache.scala
index 8ed0091f..742905b7 100644
--- a/rocket/src/main/scala/icache.scala
+++ b/rocket/src/main/scala/icache.scala
@@ -5,7 +5,7 @@ import uncore._
 import Util._
 
 case class ICacheConfig(sets: Int, assoc: Int,
-                        ibytes: Int = 4, rowbytes: Int = 64,
+                        ibytes: Int = 4, rowbytes: Int = 16,
                         ntlb: Int = 8, 
                         tl: TileLinkConfiguration,
                         as: AddressSpaceConfiguration,
@@ -187,10 +187,25 @@ class ICache(implicit c: ICacheConfig) extends Module
   val s2_miss = s2_valid && !s2_any_tag_hit
   rdy := state === s_ready && !s2_miss
 
-  //assert(!co.isVoluntary(io.mem.grant.bits.payload) || !io.mem.grant.valid, "UncachedRequestors shouldn't get voluntary grants.")
-  val (rf_cnt, refill_done) = (if(c.refillcycles > 1) Counter(io.mem.grant.valid, c.refillcycles) else (UInt(0), state === s_refill))
-  val repl_way = if (c.dm) UInt(0) else LFSR16(s2_miss)(log2Up(c.assoc)-1,0)
+  var refill_cnt = UInt(0)
+  var refill_done = state === s_refill 
+  var refill_valid = io.mem.grant.valid
+  var refill_bits = io.mem.grant.bits
+  def doRefill(g: Grant): Bool = Bool(true)
+  if(c.refillcycles > 1) {
+    val ser = Module(new FlowThroughSerializer(io.mem.grant.bits, c.refillcycles, doRefill))
+    ser.io.in <> io.mem.grant
+    refill_cnt = ser.io.cnt
+    refill_done = ser.io.done
+    refill_valid = ser.io.out.valid
+    refill_bits = ser.io.out.bits
+    ser.io.out.ready := Bool(true)
+  } else {
+    io.mem.grant.ready := Bool(true)
+  }
+  //assert(!c.tlco.isVoluntary(refill_bits.payload) || !refill_valid, "UncachedRequestors shouldn't get voluntary grants.")
 
+  val repl_way = if (c.dm) UInt(0) else LFSR16(s2_miss)(log2Up(c.assoc)-1,0)
   val enc_tagbits = c.code.width(c.tagbits)
   val tag_array = Mem(Bits(width = enc_tagbits*c.assoc), c.sets, seqRead = true)
   val tag_raddr = Reg(UInt())
@@ -240,14 +255,14 @@ class ICache(implicit c: ICacheConfig) extends Module
   for (i <- 0 until c.assoc) {
     val data_array = Mem(Bits(width = c.code.width(c.rowbits)), c.sets*c.refillcycles, seqRead = true)
     val s1_raddr = Reg(UInt())
-    when (io.mem.grant.valid && repl_way === UInt(i)) {
-      val d = io.mem.grant.bits.payload.data
-      if(c.refillcycles > 1) data_array(Cat(s2_idx,rf_cnt)) := c.code.encode(d)
-      else                   data_array(s2_idx) := c.code.encode(d)
+    when (refill_valid && repl_way === UInt(i)) {
+      val e_d = c.code.encode(refill_bits.payload.data)
+      if(c.refillcycles > 1) data_array(Cat(s2_idx,refill_cnt)) := e_d
+      else                   data_array(s2_idx) := e_d
     }
 //    /*.else*/when (s0_valid) { // uncomment ".else" to infer 6T SRAM
     .elsewhen (s0_valid) {
-      s1_raddr := s0_pgoff(c.untagbits-1,c.offbits-(if(c.refillcycles > 1) rf_cnt.getWidth else 0))
+      s1_raddr := s0_pgoff(c.untagbits-1,c.offbits-(if(c.refillcycles > 1) refill_cnt.getWidth else 0))
     }
     // if s1_tag_match is critical, replace with partial tag check
     when (s1_valid && rdy && !stall && (Bool(c.dm) || s1_tag_match(i))) { s2_dout(i) := data_array(s1_raddr) }
@@ -257,16 +272,15 @@ class ICache(implicit c: ICacheConfig) extends Module
   io.resp.bits.datablock := Mux1H(s2_tag_hit, s2_dout)
 
   val ack_q = Module(new Queue(new LogicalNetworkIO(new GrantAck), 1))
-  ack_q.io.enq.valid := refill_done && tl.co.requiresAckForGrant(io.mem.grant.bits.payload.g_type)
-  ack_q.io.enq.bits.payload.master_xact_id := io.mem.grant.bits.payload.master_xact_id
-  ack_q.io.enq.bits.header.dst := io.mem.grant.bits.header.src
+  ack_q.io.enq.valid := refill_done && tl.co.requiresAckForGrant(refill_bits.payload.g_type)
+  ack_q.io.enq.bits.payload.master_xact_id := refill_bits.payload.master_xact_id
+  ack_q.io.enq.bits.header.dst := refill_bits.header.src
 
   // output signals
   io.resp.valid := s2_hit
   io.mem.acquire.valid := (state === s_request) && ack_q.io.enq.ready
   io.mem.acquire.bits.payload := Acquire(tl.co.getUncachedReadAcquireType, s2_addr >> UInt(c.offbits), UInt(0))
   io.mem.grant_ack <> ack_q.io.deq
-  io.mem.grant.ready := Bool(true)
 
   // control state machine
   switch (state) {
diff --git a/rocket/src/main/scala/nbdcache.scala b/rocket/src/main/scala/nbdcache.scala
index 39390aae..a4b686e5 100644
--- a/rocket/src/main/scala/nbdcache.scala
+++ b/rocket/src/main/scala/nbdcache.scala
@@ -9,13 +9,10 @@ case class DCacheConfig(sets: Int, ways: Int,
                         tl: TileLinkConfiguration,
                         as: AddressSpaceConfiguration,
                         reqtagbits: Int, databits: Int,
-                        rowwords: Int = 8,
+                        rowwords: Int = 2,
                         code: Code = new IdentityCode,
                         narrowRead: Boolean = true)
 {
-  require(states > 0)
-  require(isPow2(sets))
-  require(isPow2(ways)) // TODO: relax this
   def states = tl.co.nClientStates
   def lines = sets*ways
   def dm = ways == 1
@@ -35,7 +32,7 @@ case class DCacheConfig(sets: Int, ways: Int,
   def rowbits = rowwords*databits
   def rowbytes = rowwords*databytes
   def rowoffbits = log2Up(rowbytes)
-  def refillcycles = tl.dataBits/(rowwords*databits)
+  def refillcycles = tl.dataBits/(rowbits)
   def isNarrowRead = narrowRead && databits*ways % rowbits == 0
   val statebits = log2Up(states)
   val metabits = statebits + tagbits
@@ -43,6 +40,11 @@ case class DCacheConfig(sets: Int, ways: Int,
   val encmetabits = code.width(metabits)
   val encrowbits = rowwords*encdatabits
   val lrsc_cycles = 32 // ISA requires 16-insn LRSC sequences to succeed
+
+  require(states > 0)
+  require(isPow2(sets))
+  require(isPow2(ways)) // TODO: relax this
+  require(rowbits <= tl.dataBits)
 }
 
 abstract trait DCacheBundle extends Bundle {
@@ -200,6 +202,7 @@ class MSHR(id: Int)(implicit conf: DCacheConfig) extends Module {
   val idx_match = req_idx === io.req_bits.addr(conf.untagbits-1,conf.offbits)
   val sec_rdy = idx_match && (state === s_wb_req || state === s_wb_resp || state === s_meta_clear || (state === s_refill_req || state === s_refill_resp) && !tl.co.needsTransactionOnSecondaryMiss(req_cmd, io.mem_req.bits))
 
+  require(isPow2(conf.refillcycles))
   val reply = io.mem_grant.valid && io.mem_grant.bits.payload.client_xact_id === UInt(id)
   val refill_done = reply && (if(conf.refillcycles > 1) refill_count.andR else Bool(true))
   val wb_done = reply && (state === s_wb_resp)
@@ -302,7 +305,6 @@ class MSHR(id: Int)(implicit conf: DCacheConfig) extends Module {
   io.mem_req.bits.addr := Cat(io.tag, req_idx).toUInt
   io.mem_req.bits.client_xact_id := Bits(id)
   io.mem_finish <> ackq.io.deq
-  io.mem_req.bits.client_xact_id := Bits(id)
 
   io.meta_read.valid := state === s_drain_rpq
   io.meta_read.bits.addr := io.mem_req.bits.addr << conf.offbits
@@ -430,59 +432,73 @@ class WritebackUnit(implicit conf: DCacheConfig) extends Module {
     val release = Decoupled(new Release)
   }
 
-  require(conf.refillcycles == 1) // TODO Currently will issue refillcycles distinct releases; need to merge if rowsize < tilelink.dataSize
-
-  val valid = Reg(init=Bool(false))
+  val active = Reg(init=Bool(false))
   val r1_data_req_fired = Reg(init=Bool(false))
   val r2_data_req_fired = Reg(init=Bool(false))
-  val cmd_sent = Reg(Bool())
-  val cnt = Reg(UInt(width = log2Up(conf.refillcycles+1)))
+  val cnt = Reg(init = UInt(0, width = log2Up(conf.refillcycles+1)))
   val req = Reg(new WritebackReq)
 
-  when (valid) {
+  io.release.valid := false
+  when (active) {
     r1_data_req_fired := false
     r2_data_req_fired := r1_data_req_fired
     when (io.data_req.fire() && io.meta_read.fire()) {
       r1_data_req_fired := true
       cnt := cnt + 1
     }
-    when (r2_data_req_fired && !io.release.ready) {
-      r1_data_req_fired := false
-      r2_data_req_fired := false
-      cnt := (if(conf.refillcycles > 1) cnt - Mux[UInt](r1_data_req_fired, 2, 1) else UInt(0))
+    if(conf.refillcycles > 1) { // Coalescing buffer inserted
+      when (!r1_data_req_fired && !r2_data_req_fired && cnt === conf.refillcycles) {
+        io.release.valid := true
+        active := !io.release.ready
+      }
+    } else {                    // No buffer, data released a cycle earlier
+      when (r2_data_req_fired) {
+        io.release.valid := true
+        when(!io.release.ready) {
+          r1_data_req_fired := false
+          r2_data_req_fired := false
+          cnt := UInt(0)
+        } .otherwise {
+          active := false
+        }
+      }
     }
-    when (io.release.fire()) {
-      cmd_sent := true
-    }
-    when (!r1_data_req_fired && !r2_data_req_fired && cmd_sent && cnt === conf.refillcycles) {
-      valid := false
-    }
-
   }
   when (io.req.fire()) {
-    valid := true
-    cmd_sent := false
+    active := true
     cnt := 0
     req := io.req.bits
   }
 
-  val fire = valid && cnt < UInt(conf.refillcycles)
-  io.req.ready := !valid
-  io.data_req.valid := fire
-  io.data_req.bits.way_en := req.way_en
-  io.data_req.bits.addr := (if(conf.refillcycles > 1) Cat(req.idx, cnt(log2Up(conf.refillcycles)-1,0)) 
-                            else req.idx) << conf.rowoffbits
-
-  io.release.valid := valid && r2_data_req_fired
-  io.release.bits.r_type := req.r_type
-  io.release.bits.addr := Cat(req.tag, req.idx).toUInt
-  io.release.bits.client_xact_id := req.client_xact_id
-  io.release.bits.master_xact_id := req.master_xact_id
-  io.release.bits.data := io.data_resp
+  val fire = active && cnt < UInt(conf.refillcycles)
+  io.req.ready := !active
 
   // We reissue the meta read as it sets up the muxing for s2_data_muxed
   io.meta_read.valid := fire
   io.meta_read.bits.addr := io.release.bits.addr << conf.offbits
+
+  io.data_req.valid := fire
+  io.data_req.bits.way_en := req.way_en
+  if(conf.refillcycles > 1) {
+    io.data_req.bits.addr := Cat(req.idx, cnt(log2Up(conf.refillcycles)-1,0)) << conf.rowoffbits
+  } else {
+    io.data_req.bits.addr := req.idx << conf.rowoffbits
+  }
+
+  io.release.bits.r_type := req.r_type
+  io.release.bits.addr := Cat(req.tag, req.idx).toUInt
+  io.release.bits.client_xact_id := req.client_xact_id
+  io.release.bits.master_xact_id := req.master_xact_id
+  if(conf.refillcycles > 1) {
+    val data_buf = Reg(Bits())
+    when(active && r2_data_req_fired) {
+      data_buf := Cat(io.data_resp, data_buf(conf.refillcycles*conf.encrowbits-1, conf.encrowbits))
+    }
+    io.release.bits.data := data_buf
+  } else {
+    io.release.bits.data := io.data_resp
+  }
+
 }
 
 class ProbeUnit(implicit conf: DCacheConfig) extends Module {
@@ -541,7 +557,7 @@ class ProbeUnit(implicit conf: DCacheConfig) extends Module {
   }
 
   io.req.ready := state === s_invalid
-  io.rep.valid := state === s_release && !tl.co.needsWriteback(line_state)
+  io.rep.valid := state === s_release && !(hit && tl.co.needsWriteback(line_state))
   io.rep.bits := Release(tl.co.getReleaseTypeOnProbe(req, Mux(hit, line_state, tl.co.newStateOnFlush)), req.addr, req.client_xact_id, req.master_xact_id)
 
   io.meta_read.valid := state === s_meta_read
@@ -827,8 +843,6 @@ class HellaCache(implicit conf: DCacheConfig) extends Module {
   // data
   val data = Module(new DataArray)
   val readArb = Module(new Arbiter(new DataReadReq, 4))
-  readArb.io.out.ready := !io.mem.grant.valid || io.mem.grant.ready // insert bubble if refill gets blocked
-  readArb.io.out <> data.io.read
 
   val writeArb = Module(new Arbiter(new DataWriteReq, 2))
   data.io.write.valid := writeArb.io.out.valid
@@ -912,7 +926,9 @@ class HellaCache(implicit conf: DCacheConfig) extends Module {
   }
 
   writeArb.io.in(0).bits.addr := s3_req.addr
-  writeArb.io.in(0).bits.wmask := UInt(1) << s3_req.addr(conf.rowoffbits-1,offsetlsb).toUInt
+  writeArb.io.in(0).bits.wmask := UInt(1) << (if(conf.rowoffbits > offsetlsb) 
+                                                s3_req.addr(conf.rowoffbits-1,offsetlsb).toUInt
+                                              else UInt(0))
   writeArb.io.in(0).bits.data := Fill(conf.rowwords, s3_req.data)
   writeArb.io.in(0).valid := s3_valid
   writeArb.io.in(0).bits.way_en :=  s3_way
@@ -932,8 +948,6 @@ class HellaCache(implicit conf: DCacheConfig) extends Module {
   mshrs.io.req.bits.data := s2_req.data
   when (mshrs.io.req.fire()) { replacer.miss }
 
-  mshrs.io.mem_grant.valid := io.mem.grant.fire()
-  mshrs.io.mem_grant.bits := io.mem.grant.bits
   io.mem.acquire <> DecoupledLogicalNetworkIOWrapper(mshrs.io.mem_req)
 
   // replays
@@ -944,6 +958,7 @@ class HellaCache(implicit conf: DCacheConfig) extends Module {
   s1_replay := mshrs.io.replay.valid && readArb.io.in(1).ready
   metaReadArb.io.in(1) <> mshrs.io.meta_read
   metaWriteArb.io.in(0) <> mshrs.io.meta_write
+
   // probes
   val releaseArb = Module(new Arbiter(new Release, 2))
   DecoupledLogicalNetworkIOWrapper(releaseArb.io.out) <> io.mem.release
@@ -960,12 +975,21 @@ class HellaCache(implicit conf: DCacheConfig) extends Module {
   prober.io.mshr_rdy := mshrs.io.probe_rdy
 
   // refills
-  val refill = tl.co.messageUpdatesDataArray(io.mem.grant.bits.payload)
-  writeArb.io.in(1).valid := io.mem.grant.valid && refill
-  io.mem.grant.ready := writeArb.io.in(1).ready || !refill
+  def doRefill(g: Grant): Bool = tl.co.messageUpdatesDataArray(g)
+  val refill = if(conf.refillcycles > 1) {
+    val ser = Module(new FlowThroughSerializer(io.mem.grant.bits, conf.refillcycles, doRefill))
+    ser.io.in <> io.mem.grant
+    ser.io.out
+  } else io.mem.grant
+  mshrs.io.mem_grant.valid := refill.fire()
+  mshrs.io.mem_grant.bits := refill.bits
+  refill.ready := writeArb.io.in(1).ready || !doRefill(refill.bits.payload)
+  writeArb.io.in(1).valid := refill.valid && doRefill(refill.bits.payload)
   writeArb.io.in(1).bits := mshrs.io.mem_resp
   writeArb.io.in(1).bits.wmask := SInt(-1)
-  writeArb.io.in(1).bits.data := io.mem.grant.bits.payload.data
+  writeArb.io.in(1).bits.data := refill.bits.payload.data(conf.encrowbits-1,0)
+  readArb.io.out.ready := !refill.valid || refill.ready // insert bubble if refill gets blocked
+  readArb.io.out <> data.io.read
 
   // writebacks
   val wbArb = Module(new Arbiter(new WritebackReq, 2))
diff --git a/rocket/src/main/scala/util.scala b/rocket/src/main/scala/util.scala
index 24016763..e912bb1b 100644
--- a/rocket/src/main/scala/util.scala
+++ b/rocket/src/main/scala/util.scala
@@ -1,6 +1,7 @@
 package rocket
 
 import Chisel._
+import uncore._
 import scala.math._
 
 class BooleanToInt(x: Int) {
@@ -161,3 +162,49 @@ object Random
   private def partition(value: UInt, slices: Int) =
     Vec.tabulate(slices)(i => value < round((i << value.getWidth).toDouble / slices))
 }
+
+class FlowThroughSerializer[T <: HasTileLinkData](gen: LogicalNetworkIO[T], n: Int, doSer: T => Bool) extends Module {
+  val io = new Bundle {
+    val in = Decoupled(gen.clone).flip
+    val out = Decoupled(gen.clone)
+    val cnt = UInt(OUTPUT, log2Up(n))
+    val done = Bool(OUTPUT)
+  }
+  require(io.in.bits.payload.data.width % n == 0)
+  val narrowWidth = io.in.bits.payload.data.width / n
+  val cnt = Reg(init=UInt(0, width = log2Up(n)))
+  val wrap = cnt === UInt(n-1)
+  val rbits = Reg(init=io.in.bits)
+  val active = Reg(init=Bool(false))
+
+  val shifter = Vec.fill(n){Bits(width = narrowWidth)}
+  (0 until n).foreach { 
+    i => shifter(i) := rbits.payload.data((i+1)*narrowWidth-1,i*narrowWidth)
+  }
+
+  io.done := Bool(false)
+  io.cnt := cnt
+  io.in.ready := !active
+  io.out.valid := active || io.in.valid
+  io.out.bits := io.in.bits
+  when(!active && io.in.valid) {
+    when(doSer(io.in.bits.payload)) {
+      cnt := Mux(io.out.ready, UInt(1), UInt(0))
+      rbits := io.in.bits
+      active := Bool(true)
+    }
+    io.done := !doSer(io.in.bits.payload)
+  }
+  when(active) {
+    io.out.bits := rbits
+    io.out.bits.payload.data := shifter(cnt)
+    when(io.out.ready) { 
+      cnt := cnt + UInt(1)
+      when(wrap) {
+        io.done := Bool(true)
+        active := Bool(false)
+      }
+    }
+  }
+}
+