diff --git a/multi_serdes_loopback.py b/multi_serdes_loopback.py
index c82c080..86b90e0 100644
--- a/multi_serdes_loopback.py
+++ b/multi_serdes_loopback.py
@@ -1,15 +1,15 @@
 from migen import *
-from sync_serdes import SyncSingleRX, SingleLineTX
+from sync_serdes import MultiLineRX, MultiLineTX
 from migen.genlib.fifo import SyncFIFO
 from migen.build.platforms.sinara import kasli
 from kasli_crg import KasliCRG
 from eem_helpers import generate_pads
 from uart import UART
-from io_loopback import SingleIOLoopback
+from io_loopback import SingleIOLoopback, IOLoopBack
 
 
 class MultiSerDesLoopBack(Module):
-    def __init__(self, io_pad, sys_clk_freq):
+    def __init__(self, io_pads, sys_clk_freq):
         self.uart_rx = Signal()
         self.uart_tx = Signal()
 
@@ -19,23 +19,25 @@ class MultiSerDesLoopBack(Module):
             self.uart_tx.eq(self.uart.phy_tx),
         ]
 
-        self.submodules.tx = SingleLineTX()
-        self.submodules.sync_rx = SyncSingleRX()
+        self.submodules.tx = MultiLineTX()
+        self.submodules.rx = MultiLineRX()
 
         # The actual channel
-        self.submodules.channel = SingleIOLoopback(io_pad)
+        self.submodules.channel = IOLoopBack(io_pads)
 
         # Attach FIFO to UART TX, send rate is too slow w.r.t sysclk
         self.submodules.tx_fifo = SyncFIFO(8, 64)
 
         self.comb += [
             # Repetitively send 0b00100
-            self.tx.txdata.eq(0b00100),
+            # Note: Replicate() doesn't work, 0b00100 is lowered into 3'd4
+            # I need 5'd4 for the replicate operator in Verilog
+            self.tx.txdata.eq(0b00100001000010000100),
 
             # Loopback channel
             self.channel.i.eq(self.tx.ser_out),
-            self.sync_rx.ser_in_no_dly.eq(self.channel.o),
             self.channel.t.eq(self.tx.t_out),
+            self.rx.ser_in_no_dly.eq(self.channel.o),
 
             # TX path
             self.uart.tx_data.eq(self.tx_fifo.dout),
@@ -46,30 +48,46 @@ class MultiSerDesLoopBack(Module):
         self.submodules.fsm = FSM(reset_state="WAIT_SELF_ALIGN")
 
         self.fsm.act("WAIT_SELF_ALIGN",
-            If(self.sync_rx.align_done,
+            If(self.rx.align_done,
                 NextState("SAMPLE_RXDATA"),
             ),
         )
 
-        sampled_rxdata = Signal(5)
+        sampled_rxdata = Signal(20)
 
         self.fsm.act("SAMPLE_RXDATA",
-            NextValue(sampled_rxdata, self.sync_rx.rxdata),
-            NextState("WRITE_PATTERN_UPPER"),
+            NextValue(sampled_rxdata, self.rx.rxdata),
+            NextState("WRITE_PATTERN_FIRST_UPPER"),
         )
 
-        self.fsm.act("WRITE_PATTERN_UPPER",
+        self.fsm.act("WRITE_PATTERN_FIRST_UPPER",
             If(self.tx_fifo.writable,
                 self.tx_fifo.we.eq(1),
-                self.tx_fifo.din.eq(sampled_rxdata),
-                NextState("WRITE_PATTERN_LOWER"),
+                self.tx_fifo.din.eq(sampled_rxdata[8:10]),
+                NextState("WRITE_PATTERN_FIRST_LOWER"),
             ),
         )
 
-        self.fsm.act("WRITE_PATTERN_LOWER",
+        self.fsm.act("WRITE_PATTERN_FIRST_LOWER",
             If(self.tx_fifo.writable,
                 self.tx_fifo.we.eq(1),
-                self.tx_fifo.din.eq(sampled_rxdata),
+                self.tx_fifo.din.eq(sampled_rxdata[:8]),
+                NextState("WRITE_PATTERN_SECOND_UPPER"),
+            ),
+        )
+
+        self.fsm.act("WRITE_PATTERN_SECOND_UPPER",
+            If(self.tx_fifo.writable,
+                self.tx_fifo.we.eq(1),
+                self.tx_fifo.din.eq(sampled_rxdata[18:20]),
+                NextState("WRITE_PATTERN_SECOND_LOWER"),
+            ),
+        )
+
+        self.fsm.act("WRITE_PATTERN_SECOND_LOWER",
+            If(self.tx_fifo.writable,
+                self.tx_fifo.we.eq(1),
+                self.tx_fifo.din.eq(sampled_rxdata[10:18]),
                 NextState("TERMINATE"),
             ),
         )
@@ -85,13 +103,13 @@ if __name__ == "__main__":
     # Generate pads for the I/O blocks
     eem = 3
     generate_pads(platform, eem)
-    # pads = [
-    #     platform.request("dio{}".format(eem), i) for i in range(4)
-    # ]
-    pad = platform.request("dio{}".format(eem), 0)
+    pads = [
+        platform.request("dio{}".format(eem), i) for i in range(4)
+    ]
+    # pad = platform.request("dio{}".format(eem), 0)
 
     crg = KasliCRG(platform)
-    top = MultiSerDesLoopBack(pad, crg.sys_clk_freq)
+    top = MultiSerDesLoopBack(pads, crg.sys_clk_freq)
 
     # Wire up UART core to the pads
     uart_pads = platform.request("serial")
diff --git a/sync_serdes.py b/sync_serdes.py
index 1b88233..a5d4cfc 100644
--- a/sync_serdes.py
+++ b/sync_serdes.py
@@ -640,3 +640,50 @@ class SyncSingleRX(Module):
             self.align_done.eq(1),
             NextState("INTRA_ALIGN_DONE"),
         )
+
+
+class MultiLineTX(Module):
+    def __init__(self):
+        # Ports
+        # IN: Unserialized data
+        self.txdata = Signal(20)
+        # OUT: Serialized data
+        self.ser_out = Signal(4)
+        # OUT: 3-state signal output
+        self.t_out = Signal(4)
+
+        for idx in range(4):
+            single_tx = SingleLineTX()
+
+            self.comb += [
+                self.ser_out[idx].eq(single_tx.ser_out),
+                self.t_out[idx].eq(single_tx.t_out),
+                single_tx.txdata.eq(self.txdata[5*idx:5*(idx+1)]),
+            ]
+
+            self.submodules += single_tx
+
+
+class MultiLineRX(Module):
+    def __init__(self):
+        # Ports
+        # IN: Undelayed serial signal
+        self.ser_in_no_dly = Signal(4)
+        # OUT: Received data after self-alignment, decimation
+        self.rxdata = Signal(20)
+        # OUT: RXDATA from all channels are self-aligned
+        self.align_done = Signal()
+
+        channel_align_done = Signal(4)
+        self.comb += self.align_done.eq(channel_align_done == 0b1111)
+
+        for idx in range(4):
+            single_rx = SyncSingleRX()
+
+            self.comb += [
+                single_rx.ser_in_no_dly.eq(self.ser_in_no_dly[idx]),
+                self.rxdata[5*idx:5*(idx+1)].eq(single_rx.rxdata),
+                channel_align_done[idx].eq(single_rx.align_done),
+            ]
+
+            self.submodules += single_rx