diff --git a/rocket/src/main/scala/icache.scala b/rocket/src/main/scala/icache.scala
index 3f627f77..5318ce7a 100644
--- a/rocket/src/main/scala/icache.scala
+++ b/rocket/src/main/scala/icache.scala
@@ -92,32 +92,32 @@ class rocketICache(sets: Int, assoc: Int, co: CoherencePolicyWithUncached) exten
     Mux((state === s_refill_wait) || (state === s_refill),  Cat(r_cpu_req_idx(indexmsb,offsetbits), refill_count),
       io.cpu.req_idx(indexmsb, offsetbits-rf_cnt_bits)).toUFix;
 
+  val tag_array = Mem(sets, seqRead = true) { Bits(width = tagbits*assoc) }
+  val tag_rdata = Reg() { Bits() }
+  when (tag_we) {
+    tag_array.write(tag_addr, Fill(assoc, r_cpu_miss_tag), FillInterleaved(tagbits, if (assoc > 1) UFixToOH(repl_way) else Bits(1)))
+  }.otherwise {
+    tag_rdata := tag_array(tag_addr)
+  }
+
+  val vb_array = Reg(resetVal = Bits(0, lines))
+  when (io.cpu.invalidate) {
+    vb_array := Bits(0)
+  }.elsewhen (tag_we) {
+    vb_array := vb_array.bitSet(Cat(r_cpu_req_idx(indexmsb,indexlsb), if (assoc > 1) repl_way else null), !invalidated)
+  }
+
   val data_mux = (new Mux1H(assoc)){Bits(width = databits)}
   var any_hit = Bool(false)
   for (i <- 0 until assoc)
   {
-    val repl_me = (repl_way === UFix(i))
-    val tag_array = Mem(sets, seqRead = true){ Bits(width = tagbits) }
-    val tag_rdata = Reg() { Bits(width = tagbits) }
-    when (tag_we && repl_me) { tag_array(tag_addr) := r_cpu_miss_tag }
-    .otherwise { tag_rdata := tag_array(tag_addr) }
-
-    // valid bit array
-    val vb_array = Reg(resetVal = Bits(0, sets));
-    when (io.cpu.invalidate) {
-      vb_array := Bits(0)
-    }
-    .elsewhen (tag_we && repl_me) {
-      vb_array := vb_array.bitSet(r_cpu_req_idx(indexmsb,indexlsb).toUFix, !invalidated)
-    }
-
-    val valid = vb_array(r_cpu_req_idx(indexmsb,indexlsb)).toBool;
-    val hit = valid && (tag_rdata === r_cpu_hit_addr(tagmsb,taglsb))
+    val valid = vb_array(Cat(r_cpu_req_idx(indexmsb,indexlsb), if (assoc > 1) UFix(i, log2Up(assoc)) else null))
+    val hit = valid && tag_rdata(tagbits*(i+1)-1, tagbits*i) === r_cpu_hit_addr(tagmsb,taglsb)
     
     // data array
     val data_array = Mem(sets*REFILL_CYCLES, seqRead = true){ io.mem.xact_rep.bits.data.clone }
     val data_out = Reg(){ io.mem.xact_rep.bits.data.clone }
-    when (io.mem.xact_rep.valid && repl_me) { data_array(data_addr) := io.mem.xact_rep.bits.data }
+    when (io.mem.xact_rep.valid && repl_way === UFix(i)) { data_array(data_addr) := io.mem.xact_rep.bits.data }
     .otherwise { data_out := data_array(data_addr) }
 
     data_mux.io.sel(i) := hit
diff --git a/rocket/src/main/scala/nbdcache.scala b/rocket/src/main/scala/nbdcache.scala
index 1a0977c8..fdc7376a 100644
--- a/rocket/src/main/scala/nbdcache.scala
+++ b/rocket/src/main/scala/nbdcache.scala
@@ -595,35 +595,6 @@ class FlushUnit(lines: Int, co: CoherencePolicy) extends Component {
   io.meta_req.bits.data.tag := UFix(0)
 }
 
-class MetaDataArray(lines: Int) extends Component {
-  val io = new Bundle {
-    val req  = (new FIFOIO) { new MetaArrayReq() }.flip
-    val resp = (new MetaData).asOutput()
-    val state_req = (new FIFOIO) { new MetaArrayReq() }.flip
-  }
-
-  val permissions_array = Mem(lines){ UFix(width = 2) }
-  val raddr = Reg() { Bits() }
-  when (io.state_req.valid && io.state_req.bits.rw) {
-    permissions_array(io.state_req.bits.idx) := io.state_req.bits.data.state
-  }
-  when (io.req.valid) {
-    when (io.req.bits.rw) { permissions_array(io.req.bits.idx) := io.req.bits.data.state }
-    .otherwise { raddr := io.req.bits.idx }
-  }
-
-  val tag_array = Mem(lines, seqRead = true){ Bits(width=TAG_BITS) }
-  val tag_rdata = Reg() { Bits() }
-  when (io.req.valid) {
-    when (io.req.bits.rw) { tag_array(io.req.bits.idx) := io.req.bits.data.tag }
-    .otherwise { tag_rdata := tag_array(io.req.bits.idx) }
-  }
-
-  io.resp.state := permissions_array(raddr)
-  io.resp.tag   := tag_rdata
-  io.req.ready  := Bool(true)
-}
-
 class MetaDataArrayArray(lines: Int) extends Component {
   val io = new Bundle {
     val req  = (new FIFOIO) { new MetaArrayReq() }.flip
@@ -632,18 +603,32 @@ class MetaDataArrayArray(lines: Int) extends Component {
     val way_en = Bits(width = NWAYS, dir = OUTPUT)
   }
 
+  val permBits = io.req.bits.data.state.width
+  val perms = Mem(lines) { UFix(width = permBits*NWAYS) }
+  val tags = Mem(lines*NWAYS, seqRead = true) { Bits(width = TAG_BITS*NWAYS) }
+  val tag = Reg() { Bits() }
+  val raddr = Reg() { Bits() }
   val way_en_ = Reg { Bits(width=NWAYS) }
-  when (io.req.valid && io.req.ready) {
+
+  when (io.state_req.valid && io.state_req.bits.rw) {
+    perms.write(io.state_req.bits.idx, Fill(NWAYS, io.state_req.bits.data.state), FillInterleaved(permBits, io.state_req.bits.way_en))
+  }
+  when (io.req.valid) {
+    when (io.req.bits.rw) {
+      perms.write(io.req.bits.idx, Fill(NWAYS, io.req.bits.data.state), FillInterleaved(permBits, io.req.bits.way_en))
+      tags.write(io.req.bits.idx, Fill(NWAYS, io.req.bits.data.tag), FillInterleaved(TAG_BITS, io.req.bits.way_en))
+    }
+    .otherwise {
+      raddr := io.req.bits.idx
+      tag := tags(io.req.bits.idx)
+    }
     way_en_ := io.req.bits.way_en
   }
 
+  val perm = perms(raddr)
   for(w <- 0 until NWAYS) {
-    val way = new MetaDataArray(lines)
-    way.io.req.bits <> io.req.bits
-    way.io.req.valid := io.req.valid && io.req.bits.way_en(w).toBool
-    way.io.state_req.bits <> io.state_req.bits
-    way.io.state_req.valid := io.state_req.valid && io.state_req.bits.way_en(w).toBool
-    way.io.resp <> io.resp(w)
+    io.resp(w).state := perm(permBits*(w+1)-1, permBits*w)
+    io.resp(w).tag := tag(TAG_BITS*(w+1)-1, TAG_BITS*w)
   }
 
   io.way_en := way_en_