diff --git a/artiq/coredevice/rtio.py b/artiq/coredevice/rtio.py
index e28f06e06..ae7a5ee03 100644
--- a/artiq/coredevice/rtio.py
+++ b/artiq/coredevice/rtio.py
@@ -21,17 +21,17 @@ class LLRTIOOut(AutoDB):
 
     @kernel
     def _set_oe(self):
-        syscall("rtio_oe", self.channel, True)
+        syscall("rtio_set_oe", time_to_cycles(now()), self.channel, True)
 
     @kernel
-    def set_value(self, t, value):
-        """Sets the value of the RTIO channel.
+    def set_o(self, t, value):
+        """Sets the output value of the RTIO channel.
 
         :param t: timestamp in RTIO cycles (64-bit integer).
         :param value: value to set at the output.
 
         """
-        syscall("rtio_set", t, self.channel, value)
+        syscall("rtio_set_o", t, self.channel, value)
 
     @kernel
     def on(self, t):
@@ -40,7 +40,7 @@ class LLRTIOOut(AutoDB):
         :param t: timestamp in RTIO cycles (64-bit integer).
 
         """
-        self.set_value(t, 1)
+        self.set_o(t, 1)
 
     @kernel
     def off(self, t):
@@ -49,28 +49,10 @@ class LLRTIOOut(AutoDB):
         :param t: timestamp in RTIO cycles (64-bit integer).
 
         """
-        self.set_value(t, 0)
+        self.set_o(t, 0)
 
 
-class _RTIOBase(AutoDB):
-    class DBKeys:
-        core = Device()
-        channel = Argument()
-
-    def build(self):
-        self.previous_timestamp = int64(0)  # in RTIO cycles
-
-    @kernel
-    def _set_oe(self, oe):
-        syscall("rtio_oe", self.channel, oe)
-
-    @kernel
-    def _set_value(self, value):
-        syscall("rtio_set", time_to_cycles(now()), self.channel, value)
-        self.previous_timestamp = time_to_cycles(now())
-
-
-class RTIOOut(_RTIOBase):
+class RTIOOut(AutoDB):
     """RTIO output driver.
 
     Configures the corresponding RTIO channel as output on the core device and
@@ -85,9 +67,22 @@ class RTIOOut(_RTIOBase):
     :param channel: channel number
 
     """
+    class DBKeys:
+        core = Device()
+        channel = Argument()
+
     def build(self):
-        _RTIOBase.build(self)
-        self._set_oe(True)
+        self.previous_timestamp = int64(0)  # in RTIO cycles
+        self._set_oe()
+
+    @kernel
+    def _set_oe(self):
+        syscall("rtio_set_oe", time_to_cycles(now()), self.channel, True)
+
+    @kernel
+    def _set_o(self, value):
+        syscall("rtio_set_o", time_to_cycles(now()), self.channel, value)
+        self.previous_timestamp = time_to_cycles(now())
 
     @kernel
     def sync(self):
@@ -96,24 +91,19 @@ class RTIOOut(_RTIOBase):
         This function is useful to synchronize CPU-controlled devices (such as
         the AD9858 DDS bus) with related RTIO controls (such as RF switches at
         the output of the DDS).
-
         """
         while syscall("rtio_get_counter") < self.previous_timestamp:
             pass
 
     @kernel
     def on(self):
-        """Sets the output to a logic high state.
-
-        """
-        self._set_value(1)
+        """Sets the output to a logic high state."""
+        self._set_o(1)
 
     @kernel
     def off(self):
-        """Sets the output to a logic low state.
-
-        """
-        self._set_value(0)
+        """Sets the output to a logic low state."""
+        self._set_o(0)
 
     @kernel
     def pulse(self, duration):
@@ -125,7 +115,7 @@ class RTIOOut(_RTIOBase):
         self.off()
 
 
-class RTIOIn(_RTIOBase):
+class RTIOIn(AutoDB):
     """RTIO input driver.
 
     Configures the corresponding RTIO channel as input on the core device and
@@ -134,29 +124,41 @@ class RTIOIn(_RTIOBase):
 
     :param core: core device
     :param channel: channel number
-
     """
+    class DBKeys:
+        core = Device()
+        channel = Argument()
+
     def build(self):
-        _RTIOBase.build(self)
-        self._set_oe(False)
+        self.previous_timestamp = int64(0)  # in RTIO cycles
+        self._set_oe()
+
+    @kernel
+    def _set_oe(self):
+        syscall("rtio_set_oe", time_to_cycles(now()), self.channel, False)
+
+    @kernel
+    def _set_sensitivity(self, value):
+        syscall("rtio_set_sensitivity", time_to_cycles(now()), self.channel, value)
+        self.previous_timestamp = time_to_cycles(now())
 
     @kernel
     def gate_rising(self, duration):
         """Register rising edge events for the specified duration.
 
         """
-        self._set_value(1)
+        self._set_sensitivity(1)
         delay(duration)
-        self._set_value(0)
+        self._set_sensitivity(0)
 
     @kernel
     def gate_falling(self, duration):
         """Register falling edge events for the specified duration.
 
         """
-        self._set_value(2)
+        self._set_sensitivity(2)
         delay(duration)
-        self._set_value(0)
+        self._set_sensitivity(0)
 
     @kernel
     def gate_both(self, duration):
@@ -164,18 +166,9 @@ class RTIOIn(_RTIOBase):
         duration.
 
         """
-        self._set_value(3)
+        self._set_sensitivity(3)
         delay(duration)
-        self._set_value(0)
-
-    @kernel
-    def pileup_count(self):
-        """Returns the number of pileup events (a system clock cycle with too
-        many input transitions) since the last call to this function for this
-        channel (or since the last RTIO reset).
-
-        """
-        return syscall("rtio_pileup_count", self.channel)
+        self._set_sensitivity(0)
 
     @kernel
     def count(self):
diff --git a/artiq/coredevice/runtime.py b/artiq/coredevice/runtime.py
index 4c428d643..15cfff05c 100644
--- a/artiq/coredevice/runtime.py
+++ b/artiq/coredevice/runtime.py
@@ -12,11 +12,11 @@ llvm.initialize_all_targets()
 llvm.initialize_all_asmprinters()
 
 _syscalls = {
-    "rtio_oe": "ib:n",
-    "rtio_set": "Iii:n",
+    "rtio_set_o": "Iii:n",
+    "rtio_set_oe": "Iib:n",
+    "rtio_set_sensitivity": "Iii:n",
     "rtio_get_counter": "n:I",
     "rtio_get": "iI:I",
-    "rtio_pileup_count": "i:i",
     "dds_phase_clear_en": "ib:n",
     "dds_program": "Iiiiibb:n",
 }
diff --git a/artiq/gateware/rtio/__init__.py b/artiq/gateware/rtio/__init__.py
index f170e8354..9b34976c0 100644
--- a/artiq/gateware/rtio/__init__.py
+++ b/artiq/gateware/rtio/__init__.py
@@ -1,2 +1 @@
-from artiq.gateware.rtio import phy
-from artiq.gateware.rtio.core import RTIO
+from artiq.gateware.rtio.core import Channel, RTIO
diff --git a/artiq/gateware/rtio/core.py b/artiq/gateware/rtio/core.py
index a9719f8c8..c7d764920 100644
--- a/artiq/gateware/rtio/core.py
+++ b/artiq/gateware/rtio/core.py
@@ -2,12 +2,13 @@ from fractions import Fraction
 
 from migen.fhdl.std import *
 from migen.bank.description import *
+from migen.genlib.misc import optree
 from migen.genlib.record import Record
 from migen.genlib.cdc import *
 from migen.genlib.fifo import AsyncFIFO
 from migen.genlib.resetsync import AsyncResetSynchronizer
 
-from artiq.gateware.rtio.rbus import get_fine_ts_width
+from artiq.gateware.rtio import rtlink
 
 
 class _GrayCodeTransfer(Module):
@@ -35,26 +36,20 @@ class _GrayCodeTransfer(Module):
 
 
 class _RTIOCounter(Module):
-    def __init__(self, width, loopback_latency):
+    def __init__(self, width):
         self.width = width
-        # Timestamp counter in RTIO domain for outputs
-        self.o_value_rio = Signal(width)
+        # Timestamp counter in RTIO domain
+        self.value_rio = Signal(width)
         # Timestamp counter resynchronized to sys domain
-        # Lags behind o_value_rio, monotonic and glitch-free
-        self.o_value_sys = Signal(width)
-        # Timestamp counter in RTIO domain for inputs,
-        # compensated for PHY loopback latency
-        self.i_value_rio = Signal(width, reset=2**width-loopback_latency)
+        # Lags behind value_rio, monotonic and glitch-free
+        self.value_sys = Signal(width)
 
         # # #
 
-        self.sync.rio += [
-            self.o_value_rio.eq(self.o_value_rio + 1),
-            self.i_value_rio.eq(self.i_value_rio + 1)
-        ]
+        self.sync.rio += self.value_rio.eq(self.value_rio + 1),
         gt = _GrayCodeTransfer(width)
         self.submodules += gt
-        self.comb += gt.i.eq(self.o_value_rio), self.o_value_sys.eq(gt.o)
+        self.comb += gt.i.eq(self.value_rio), self.value_sys.eq(gt.o)
 
 
 # CHOOSING A GUARD TIME
@@ -78,263 +73,235 @@ class _RTIOCounter(Module):
 #    timestamp*Tio > (fifo_depth-1)*Tio + time
 # We also have (guard time reached):
 #    timestamp*Tio < time + guard_io_cycles*Tio
-# [NB: time > counter.o_value_sys*Tio]
+# [NB: time > counter.value_sys*Tio]
 # Thus we must have:
 #    guard_io_cycles > fifo_depth-1
 #
 # We can prevent overflows by choosing instead:
 #    guard_io_cycles < fifo_depth-1
 
-class _RTIOBankO(Module):
-    def __init__(self, rbus, counter, fine_ts_width, fifo_depth, guard_io_cycles):
-        self.sel = Signal(max=len(rbus))
-        # timestamp and value must be valid 1 cycle before we
-        self.timestamp = Signal(counter.width + fine_ts_width)
-        self.value = Signal(2)
+class _OutputManager(Module):
+    def __init__(self, interface, counter, fifo_depth, guard_io_cycles):
+        data_width = rtlink.get_data_width(interface)
+        address_width = rtlink.get_address_width(interface)
+        fine_ts_width = rtlink.get_fine_ts_width(interface)
+
+        ev_layout = []
+        if data_width:
+            ev_layout.append(("data", data_width))
+        if address_width:
+            ev_layout.append(("address", address_width))
+        ev_layout.append(("timestamp", counter.width + fine_ts_width))
+        # ev must be valid 1 cycle before we to account for the latency in
+        # generating replace, sequence_error and nop
+        self.ev = Record(ev_layout)
+
         self.writable = Signal()
         self.we = Signal()  # maximum throughput 1/2
-        self.underflow = Signal()  # valid 2 cycles after we
-        self.underflow_reset = Signal()
+
+        self.underflow = Signal()  # valid 1 cycle after we, pulsed
         self.sequence_error = Signal()
-        self.sequence_error_reset = Signal()
 
         # # #
 
-        signal_underflow = Signal()
-        signal_sequence_error = Signal()
-        fifos = []
-        ev_layout = [("timestamp", counter.width + fine_ts_width),
-                     ("value", 2)]
-        for n, chif in enumerate(rbus):
-            # FIFO
-            fifo = RenameClockDomains(AsyncFIFO(ev_layout, fifo_depth),
-                                      {"write": "rsys", "read": "rio"})
-            self.submodules += fifo
-            fifos.append(fifo)
+        # FIFO
+        fifo = RenameClockDomains(AsyncFIFO(ev_layout, fifo_depth),
+                                  {"write": "rsys", "read": "rio"})
+        self.submodules += fifo
 
-            # Buffer
-            buf_pending = Signal()
-            buf = Record(ev_layout)
-            buf_just_written = Signal()
+        # Buffer
+        buf_pending = Signal()
+        buf = Record(ev_layout)
+        buf_just_written = Signal()
 
-            # Special cases
-            replace = Signal()
-            sequence_error = Signal()
-            nop = Signal()
-            self.sync.rsys += [
-                replace.eq(self.timestamp == buf.timestamp[fine_ts_width:]),
-                sequence_error.eq(self.timestamp < buf.timestamp[fine_ts_width:]),
-                nop.eq(self.value == buf.value)
-            ]
-            self.comb += If(self.we & (self.sel == n) & sequence_error,
-                signal_sequence_error.eq(1))
-
-            # Buffer read and FIFO write
-            self.comb += fifo.din.eq(buf)
-            in_guard_time = Signal()
-            self.comb += in_guard_time.eq(
-                buf.timestamp[fine_ts_width:] < counter.o_value_sys + guard_io_cycles)
-            self.sync.rsys += If(in_guard_time, buf_pending.eq(0))
-            self.comb += \
-                If(buf_pending,
-                    If(in_guard_time,
-                        If(buf_just_written,
-                            signal_underflow.eq(1)
-                        ).Else(
-                            fifo.we.eq(1)
-                        )
-                    ),
-                    If((self.we & (self.sel == n)
-                            & ~replace & ~nop & ~sequence_error),
-                       fifo.we.eq(1)
-                    )
-                )
-
-            # Buffer write
-            # Must come after read to handle concurrent read+write properly
-            self.sync.rsys += [
-                buf_just_written.eq(0),
-                If(self.we & (self.sel == n) & ~nop & ~sequence_error,
-                    buf_just_written.eq(1),
-                    buf_pending.eq(1),
-                    buf.timestamp.eq(self.timestamp),
-                    buf.value.eq(self.value)
-                )
-            ]
-
-            # Buffer output of FIFO to improve timing
-            dout_stb = Signal()
-            dout_ack = Signal()
-            dout = Record(ev_layout)
-            self.sync.rio += \
-                If(fifo.re,
-                    dout_stb.eq(1),
-                    dout.eq(fifo.dout)
-                ).Elif(dout_ack,
-                    dout_stb.eq(0)
-                )
-            self.comb += fifo.re.eq(fifo.readable & (~dout_stb | dout_ack))
-
-            # FIFO read through buffer
-            self.comb += [
-                dout_ack.eq(
-                    dout.timestamp[fine_ts_width:] == counter.o_value_rio),
-                chif.o_stb.eq(dout_stb & dout_ack),
-                chif.o_value.eq(dout.value)
-            ]
-            if fine_ts_width:
-                self.comb += chif.o_fine_ts.eq(
-                    dout.timestamp[:fine_ts_width])
-
-        self.comb += \
-            self.writable.eq(Array(fifo.writable for fifo in fifos)[self.sel])
+        # Special cases
+        replace = Signal()
+        sequence_error = Signal()
+        nop = Signal()
         self.sync.rsys += [
-            If(self.underflow_reset, self.underflow.eq(0)),
-            If(self.sequence_error_reset, self.sequence_error.eq(0)),
-            If(signal_underflow, self.underflow.eq(1)),
-            If(signal_sequence_error, self.sequence_error.eq(1))
+            replace.eq(self.ev.timestamp == buf.timestamp[fine_ts_width:]),
+            sequence_error.eq(self.ev.timestamp < buf.timestamp[fine_ts_width:])
         ]
+        if interface.suppress_nop:
+            self.sync.rsys += nop.eq(
+                optree("&", 
+                       [getattr(self.ev, a) == getattr(buf, a)
+                        for a in ("data", "address")
+                        if hasattr(self.ev, a)],
+                       default=0))
+        self.comb += self.sequence_error.eq(self.we & sequence_error)
+
+        # Buffer read and FIFO write
+        self.comb += fifo.din.eq(buf)
+        in_guard_time = Signal()
+        self.comb += in_guard_time.eq(
+            buf.timestamp[fine_ts_width:]
+                < counter.value_sys + guard_io_cycles)
+        self.sync.rsys += If(in_guard_time, buf_pending.eq(0))
+        self.comb += \
+            If(buf_pending,
+                If(in_guard_time,
+                    If(buf_just_written,
+                        self.underflow.eq(1)
+                    ).Else(
+                        fifo.we.eq(1)
+                    )
+                ),
+                If(self.we & ~replace & ~nop & ~sequence_error,
+                   fifo.we.eq(1)
+                )
+            )
+
+        # Buffer write
+        # Must come after read to handle concurrent read+write properly
+        self.sync.rsys += [
+            buf_just_written.eq(0),
+            If(self.we & ~nop & ~sequence_error,
+                buf_just_written.eq(1),
+                buf_pending.eq(1),
+                buf.eq(self.ev)
+            )
+        ]
+        self.comb += self.writable.eq(fifo.writable)
+
+        # Buffer output of FIFO to improve timing
+        dout_stb = Signal()
+        dout_ack = Signal()
+        dout = Record(ev_layout)
+        self.sync.rio += \
+            If(fifo.re,
+                dout_stb.eq(1),
+                dout.eq(fifo.dout)
+            ).Elif(dout_ack,
+                dout_stb.eq(0)
+            )
+        self.comb += fifo.re.eq(fifo.readable & (~dout_stb | dout_ack))
+
+        # FIFO read through buffer
+        # TODO: report error on stb & busy
+        self.comb += [
+            dout_ack.eq(
+                dout.timestamp[fine_ts_width:] == counter.value_rio),
+            interface.stb.eq(dout_stb & dout_ack)
+        ]
+        if data_width:
+            self.comb += interface.data.eq(dout.data)
+        if address_width:
+            self.comb += interface.address.eq(dout.address)
+        if fine_ts_width:
+            self.comb += interface.fine_ts.eq(dout.timestamp[:fine_ts_width])
 
 
-class _RTIOBankI(Module):
-    def __init__(self, rbus, counter, fine_ts_width, fifo_depth):
-        self.sel = Signal(max=len(rbus))
-        self.timestamp = Signal(counter.width + fine_ts_width)
-        self.value = Signal()
+class _InputManager(Module):
+    def __init__(self, interface, counter, fifo_depth):
+        data_width = rtlink.get_data_width(interface)
+        fine_ts_width = rtlink.get_fine_ts_width(interface)
+
+        ev_layout = []
+        if data_width:
+            ev_layout.append(("data", data_width))
+        if interface.timestamped:
+            ev_layout.append(("timestamp", counter.width + fine_ts_width))
+        self.ev = Record(ev_layout)
+
         self.readable = Signal()
         self.re = Signal()
-        self.overflow = Signal()
-        self.overflow_reset = Signal()
-        self.pileup_count = Signal(16)
-        self.pileup_reset = Signal()
+        
+        self.overflow = Signal()  # pulsed
 
         # # #
 
-        timestamps = []
-        values = []
-        readables = []
-        overflows = []
-        pileup_counts = []
-        ev_layout = [("timestamp", counter.width+fine_ts_width),
-                     ("value", 1)]
-        for n, chif in enumerate(rbus):
-            if hasattr(chif, "oe"):
-                sensitivity = Signal(2)
-                self.sync.rio += If(~chif.oe & chif.o_stb,
-                                    sensitivity.eq(chif.o_value))
+        fifo = RenameClockDomains(AsyncFIFO(ev_layout, fifo_depth),
+                                  {"read": "rsys", "write": "rio"})
+        self.submodules += fifo
 
-                fifo = RenameClockDomains(AsyncFIFO(ev_layout, fifo_depth),
-                                          {"read": "rsys", "write": "rio"})
-                self.submodules += fifo
-
-                # FIFO write
-                if fine_ts_width:
-                    full_ts = Cat(chif.i_fine_ts, counter.i_value_rio)
-                else:
-                    full_ts = counter.i_value_rio
-                self.comb += [
-                    fifo.din.timestamp.eq(full_ts),
-                    fifo.din.value.eq(chif.i_value),
-                    fifo.we.eq(
-                        ~chif.oe & chif.i_stb &
-                        ((chif.i_value & sensitivity[0])
-                            | (~chif.i_value & sensitivity[1])))
-                ]
-
-                # FIFO read
-                timestamps.append(fifo.dout.timestamp)
-                values.append(fifo.dout.value)
-                readables.append(fifo.readable)
-                self.comb += fifo.re.eq(self.re & (self.sel == n))
-
-                overflow = Signal()
-                overflow_reset_sync = PulseSynchronizer("rsys", "rio")
-                self.submodules += overflow_reset_sync
-                self.comb += overflow_reset_sync.i.eq(
-                    self.overflow_reset & (self.sel == n))
-                self.sync.rio += [
-                    If(overflow_reset_sync.o, overflow.eq(0)),
-                    If(fifo.we & ~fifo.writable, overflow.eq(1))
-                ]
-                overflow_sys = Signal()
-                self.specials += MultiReg(overflow, overflow_sys, "rsys")
-                overflows.append(overflow_sys)
-
-                pileup_count = Signal(16)
-                pileup_count_reset_sync = PulseSynchronizer("rsys", "rio")
-                self.submodules += pileup_count_reset_sync
-                self.comb += pileup_count_reset_sync.i.eq(
-                    self.pileup_reset & (self.sel == n))
-                self.sync.rio += \
-                    If(pileup_count_reset_sync.o,
-                        pileup_count.eq(0)
-                    ).Elif(chif.i_pileup,
-                        If(pileup_count != 2**16 - 1,  # saturate
-                            pileup_count.eq(pileup_count + 1)
-                        )
-                    )
-                pileup_count_sync = _GrayCodeTransfer(16)
-                self.submodules += pileup_count_sync
-                self.comb += pileup_count_sync.i.eq(pileup_count)
-                pileup_counts.append(pileup_count_sync.o)
+        # FIFO write
+        if data_width:
+            self.comb += fifo.din.data.eq(interface.data)
+        if interface.timestamped:
+            if fine_ts_width:
+                full_ts = Cat(interface.fine_ts, counter.value_rio)
             else:
-                timestamps.append(0)
-                values.append(0)
-                readables.append(0)
-                overflows.append(0)
-                pileup_counts.append(0)
+                full_ts = counter.value_rio
+            self.comb += fifo.din.timestamp.eq(full_ts)
+        self.comb += fifo.we.eq(interface.stb)
 
+        # FIFO read
         self.comb += [
-            self.timestamp.eq(Array(timestamps)[self.sel]),
-            self.value.eq(Array(values)[self.sel]),
-            self.readable.eq(Array(readables)[self.sel]),
-            self.overflow.eq(Array(overflows)[self.sel]),
-            self.pileup_count.eq(Array(pileup_counts)[self.sel])
+            self.ev.eq(fifo.dout),
+            self.readable.eq(fifo.readable),
+            fifo.re.eq(self.re)
+        ]
+
+        overflow_sync = PulseSynchronizer("rio", "rsys")
+        overflow_ack_sync = PulseSynchronizer("rsys", "rio")
+        self.submodules += overflow_sync, overflow_ack_sync
+        overflow_blind = Signal()
+        self.comb += overflow_sync.i.eq(fifo.we & ~fifo.writable & ~overflow_blind)
+        self.sync.rio += [
+            If(fifo.we & ~fifo.writable, overflow_blind.eq(1)),
+            If(overflow_ack_sync.o, overflow_blind.eq(0))
+        ]
+        self.comb += [
+            overflow_ack_sync.i.eq(overflow_sync.o),
+            self.overflow.eq(overflow_sync.o)
         ]
 
 
-class RTIO(Module, AutoCSR):
-    def __init__(self, phy, clk_freq, counter_width=63,
-                 ofifo_depth=64, ififo_depth=64,
-                 guard_io_cycles=20):
-        fine_ts_width = get_fine_ts_width(phy.rbus)
+class Channel:
+    def __init__(self, interface, ofifo_depth=64, ififo_depth=64):
+        self.interface = interface
+        self.ofifo_depth = ofifo_depth
+        self.ififo_depth = ififo_depth
 
-        # Submodules
-        self.submodules.counter = _RTIOCounter(
-            counter_width, phy.loopback_latency)
-        self.submodules.bank_o = _RTIOBankO(
-            phy.rbus, self.counter, fine_ts_width, ofifo_depth, guard_io_cycles)
-        self.submodules.bank_i = _RTIOBankI(
-            phy.rbus, self.counter, fine_ts_width, ififo_depth)
+
+class _CSRs(AutoCSR):
+    def __init__(self):
+        self.frequency_i = CSRStatus(32)
+        self.frequency_fn = CSRStatus(8)
+        self.frequency_fd = CSRStatus(8)
+
+
+class _KernelCSRs(AutoCSR):
+    def __init__(self, chan_sel_width,
+                 data_width, address_width, full_ts_width):
+        self.reset = CSRStorage(reset=1)
+        self.chan_sel = CSRStorage(chan_sel_width)
+
+        self.o_data = CSRStorage(data_width)
+        self.o_address = CSRStorage(address_width)
+        self.o_timestamp = CSRStorage(full_ts_width)
+        self.o_we = CSR()
+        self.o_status = CSRStatus(3)
+        self.o_underflow_reset = CSR()
+        self.o_sequence_error_reset = CSR()
+
+        self.i_data = CSRStatus(data_width)        
+        self.i_timestamp = CSRStatus(full_ts_width)
+        self.i_re = CSR()
+        self.i_status = CSRStatus(2)
+        self.i_overflow_reset = CSR()
+
+        self.counter = CSRStatus(full_ts_width)
+        self.counter_update = CSR()
+
+
+class RTIO(Module):
+    def __init__(self, channels, clk_freq, counter_width=63,
+                 guard_io_cycles=20):
+        data_width = max(rtlink.get_data_width(c.interface)
+                         for c in channels)
+        address_width = max(rtlink.get_address_width(c.interface)
+                            for c in channels)
+        fine_ts_width = max(rtlink.get_fine_ts_width(c.interface)
+                            for c in channels)
 
         # CSRs
-        self._reset = CSRStorage(reset=1)
-        self._chan_sel = CSRStorage(flen(self.bank_o.sel))
-
-        self._oe = CSR()
-
-        self._o_timestamp = CSRStorage(counter_width + fine_ts_width)
-        self._o_value = CSRStorage(2)
-        self._o_we = CSR()
-        self._o_status = CSRStatus(3)
-        self._o_underflow_reset = CSR()
-        self._o_sequence_error_reset = CSR()
-
-        self._i_timestamp = CSRStatus(counter_width + fine_ts_width)
-        self._i_value = CSRStatus()
-        self._i_re = CSR()
-        self._i_status = CSRStatus(2)
-        self._i_overflow_reset = CSR()
-        self._i_pileup_count = CSRStatus(16)
-        self._i_pileup_reset = CSR()
-
-        self._counter = CSRStatus(counter_width + fine_ts_width)
-        self._counter_update = CSR()
-
-        self._frequency_i = CSRStatus(32)
-        self._frequency_fn = CSRStatus(8)
-        self._frequency_fd = CSRStatus(8)
-
+        self.csrs = _CSRs()
+        self.kcsrs = _KernelCSRs(bits_for(len(channels)-1),
+                                 max(data_width, 1),
+                                 max(address_width, 1),
+                                 counter_width + fine_ts_width)
 
         # Clocking/Reset
         # Create rsys and rio domains based on sys and rio
@@ -343,63 +310,113 @@ class RTIO(Module, AutoCSR):
         self.clock_domains.cd_rio = ClockDomain()
         self.comb += [
             self.cd_rsys.clk.eq(ClockSignal()),
-            self.cd_rsys.rst.eq(self._reset.storage)
+            self.cd_rsys.rst.eq(self.kcsrs.reset.storage)
         ]
         self.comb += self.cd_rio.clk.eq(ClockSignal("rtio"))
-        self.specials += AsyncResetSynchronizer(
-            self.cd_rio, self._reset.storage)
+        self.specials += AsyncResetSynchronizer(self.cd_rio,
+                                                self.kcsrs.reset.storage)
+
+        # Latency compensation
+        # TODO
+
+        # Managers
+        self.submodules.counter = _RTIOCounter(counter_width)
+
+        i_datas, i_timestamps = [], []
+        o_statuses, i_statuses = [], []
+        sel = self.kcsrs.chan_sel.storage
+        for n, channel in enumerate(channels):
+            selected = Signal()
+            self.comb += selected.eq(sel == n)
+
+            o_manager = _OutputManager(channel.interface.o, self.counter,
+                                       channel.ofifo_depth, guard_io_cycles)
+            self.submodules += o_manager
+
+            if hasattr(o_manager.ev, "data"):
+                self.comb += o_manager.ev.data.eq(
+                    self.kcsrs.o_data.storage)
+            if hasattr(o_manager.ev, "address"):
+                self.comb += o_manager.ev.address.eq(
+                    self.kcsrs.o_address.storage)
+            ts_shift = (flen(self.kcsrs.o_timestamp.storage)
+                        - flen(o_manager.ev.timestamp))
+            self.comb += o_manager.ev.timestamp.eq(
+                self.kcsrs.o_timestamp.storage[ts_shift:])
+
+            self.comb += o_manager.we.eq(selected & self.kcsrs.o_we.re)
+
+            underflow = Signal()
+            sequence_error = Signal()
+            self.sync.rsys += [
+                If(selected & self.kcsrs.o_underflow_reset.re,
+                   underflow.eq(0)),
+                If(selected & self.kcsrs.o_sequence_error_reset.re,
+                   sequence_error.eq(0)),
+                If(o_manager.underflow, underflow.eq(1)),
+                If(o_manager.sequence_error, sequence_error.eq(1))
+            ]
+            o_statuses.append(Cat(~o_manager.writable,
+                                  underflow,
+                                  sequence_error))
+
+            if channel.interface.i is not None:
+                i_manager = _InputManager(channel.interface.i, self.counter,
+                                          channel.ififo_depth)
+                self.submodules += i_manager
+
+                if hasattr(i_manager.ev, "data"):
+                    i_datas.append(i_manager.ev.data)
+                else:
+                    i_datas.append(0)
+                if channel.interface.i.timestamped:
+                    ts_shift = (flen(self.kcsrs.i_timestamp.status)
+                                - flen(i_manager.ev.timestamp))
+                    i_timestamps.append(i_manager.ev.timestamp << ts_shift)
+                else:
+                    i_timestamps.append(0)
+
+                self.comb += i_manager.re.eq(selected & self.kcsrs.i_re.re)
+
+                overflow = Signal()
+                self.sync.rsys += [
+                    If(selected & self.kcsrs.i_overflow_reset.re,
+                       overflow.eq(0)),
+                    If(i_manager.overflow,
+                       overflow.eq(1))
+                ]
+                i_statuses.append(Cat(~i_manager.readable, overflow))
 
-        # OE
-        oes = []
-        for n, chif in enumerate(phy.rbus):
-            if hasattr(chif, "oe"):
-                self.sync += \
-                    If(self._oe.re & (self._chan_sel.storage == n),
-                        chif.oe.eq(self._oe.r)
-                    )
-                oes.append(chif.oe)
             else:
-                oes.append(1)
-        self.comb += self._oe.w.eq(Array(oes)[self._chan_sel.storage])
-
-        # Output/Gate
+                i_datas.append(0)
+                i_timestamps.append(0)
+                i_statuses.append(0)
         self.comb += [
-            self.bank_o.sel.eq(self._chan_sel.storage),
-            self.bank_o.timestamp.eq(self._o_timestamp.storage),
-            self.bank_o.value.eq(self._o_value.storage),
-            self.bank_o.we.eq(self._o_we.re),
-            self._o_status.status.eq(Cat(~self.bank_o.writable,
-                                           self.bank_o.underflow,
-                                           self.bank_o.sequence_error)),
-            self.bank_o.underflow_reset.eq(self._o_underflow_reset.re),
-            self.bank_o.sequence_error_reset.eq(self._o_sequence_error_reset.re)
-        ]
-
-        # Input
-        self.comb += [
-            self.bank_i.sel.eq(self._chan_sel.storage),
-            self._i_timestamp.status.eq(self.bank_i.timestamp),
-            self._i_value.status.eq(self.bank_i.value),
-            self.bank_i.re.eq(self._i_re.re),
-            self._i_status.status.eq(Cat(~self.bank_i.readable, self.bank_i.overflow)),
-            self.bank_i.overflow_reset.eq(self._i_overflow_reset.re),
-            self._i_pileup_count.status.eq(self.bank_i.pileup_count),
-            self.bank_i.pileup_reset.eq(self._i_pileup_reset.re)
+            self.kcsrs.i_data.status.eq(Array(i_datas)[sel]),
+            self.kcsrs.i_timestamp.status.eq(Array(i_timestamps)[sel]),
+            self.kcsrs.o_status.status.eq(Array(o_statuses)[sel]),
+            self.kcsrs.i_status.status.eq(Array(i_statuses)[sel])
         ]
 
         # Counter access
         self.sync += \
-           If(self._counter_update.re,
-               self._counter.status.eq(Cat(Replicate(0, fine_ts_width),
-                                             self.counter.o_value_sys))
+           If(self.kcsrs.counter_update.re,
+               self.kcsrs.counter.status.eq(self.counter.value_sys
+                                                << fine_ts_width)
            )
 
-        # Frequency
+        # Frequency CSRs
         clk_freq = Fraction(clk_freq).limit_denominator(255)
         clk_freq_i = int(clk_freq)
         clk_freq_f = clk_freq - clk_freq_i
         self.comb += [
-            self._frequency_i.status.eq(clk_freq_i),
-            self._frequency_fn.status.eq(clk_freq_f.numerator),
-            self._frequency_fd.status.eq(clk_freq_f.denominator)
+            self.csrs.frequency_i.status.eq(clk_freq_i),
+            self.csrs.frequency_fn.status.eq(clk_freq_f.numerator),
+            self.csrs.frequency_fd.status.eq(clk_freq_f.denominator)
         ]
+
+    def get_csrs(self):
+        return self.csrs.get_csrs()
+
+    def get_kernel_csrs(self):
+        return self.kcsrs.get_csrs()
diff --git a/artiq/gateware/rtio/phy.py b/artiq/gateware/rtio/phy.py
deleted file mode 100644
index d1526adb6..000000000
--- a/artiq/gateware/rtio/phy.py
+++ /dev/null
@@ -1,30 +0,0 @@
-from migen.fhdl.std import *
-from migen.genlib.cdc import MultiReg
-
-from artiq.gateware.rtio.rbus import create_rbus
-
-
-class SimplePHY(Module):
-    def __init__(self, pads, output_only_pads=set()):
-        self.rbus = create_rbus(0, pads, output_only_pads)
-        self.loopback_latency = 3
-
-        # # #
-
-        for pad, chif in zip(pads, self.rbus):
-            o_pad = Signal()
-            self.sync.rio += If(chif.o_stb, o_pad.eq(chif.o_value))
-            if hasattr(chif, "oe"):
-                ts = TSTriple()
-                i_pad = Signal()
-                self.sync.rio += ts.oe.eq(chif.oe)
-                self.comb += ts.o.eq(o_pad)
-                self.specials += MultiReg(ts.i, i_pad, "rio"), \
-                    ts.get_tristate(pad)
-
-                i_pad_d = Signal()
-                self.sync.rio += i_pad_d.eq(i_pad)
-                self.comb += chif.i_stb.eq(i_pad ^ i_pad_d), \
-                    chif.i_value.eq(i_pad)
-            else:
-                self.comb += pad.eq(o_pad)
diff --git a/artiq/gateware/rtio/phy/__init__.py b/artiq/gateware/rtio/phy/__init__.py
new file mode 100644
index 000000000..e69de29bb
diff --git a/artiq/gateware/rtio/phy/ttl_simple.py b/artiq/gateware/rtio/phy/ttl_simple.py
new file mode 100644
index 000000000..d1985a012
--- /dev/null
+++ b/artiq/gateware/rtio/phy/ttl_simple.py
@@ -0,0 +1,46 @@
+from migen.fhdl.std import *
+from migen.genlib.cdc import MultiReg
+
+from artiq.gateware.rtio import rtlink
+
+
+class Output(Module):
+    def __init__(self, pad):
+        self.rtlink = rtlink.Interface(rtlink.OInterface(1))
+
+        # # #
+
+        self.sync.rio += If(self.rtlink.o.stb, pad.eq(self.rtlink.o.data))
+
+
+class Inout(Module):
+    def __init__(self, pad):
+        self.rtlink = rtlink.Interface(
+            rtlink.OInterface(2, 2),
+            rtlink.IInterface(1))
+
+        # # #
+        
+        ts = TSTriple()
+        self.specials += ts.get_tristate(pad)
+        sensitivity = Signal(2)
+
+        self.sync.rio += If(self.rtlink.o.stb,
+            Case(self.rtlink.o.address, {
+                0: ts.o.eq(self.rtlink.o.data[0]),
+                1: ts.oe.eq(self.rtlink.o.data[0]),
+                2: sensitivity.eq(self.rtlink.o.data)
+            }).makedefault()
+        )
+        
+        i = Signal()
+        i_d = Signal()
+        self.specials += MultiReg(ts.i, i, "rio")
+        self.sync.rio += i_d.eq(i)
+        self.comb += [
+            self.rtlink.i.stb.eq(
+                (sensitivity[0] & ( i & ~i_d)) |
+                (sensitivity[1] & (~i &  i_d))
+            ),
+            self.rtlink.i.data.eq(i)
+        ]
diff --git a/artiq/gateware/rtio/rbus.py b/artiq/gateware/rtio/rbus.py
deleted file mode 100644
index 821b3c2a6..000000000
--- a/artiq/gateware/rtio/rbus.py
+++ /dev/null
@@ -1,31 +0,0 @@
-from migen.fhdl.std import *
-from migen.genlib.record import Record
-
-
-def create_rbus(fine_ts_bits, pads, output_only_pads):
-    rbus = []
-    for pad in pads:
-        layout = [
-            ("o_stb", 1),
-            ("o_value", 2)
-        ]
-        if fine_ts_bits:
-            layout.append(("o_fine_ts", fine_ts_bits))
-        if pad not in output_only_pads:
-            layout += [
-                ("oe", 1),
-                ("i_stb", 1),
-                ("i_value", 1),
-                ("i_pileup", 1)
-            ]
-            if fine_ts_bits:
-                layout.append(("i_fine_ts", fine_ts_bits))
-        rbus.append(Record(layout))
-    return rbus
-
-
-def get_fine_ts_width(rbus):
-    if hasattr(rbus[0], "o_fine_ts"):
-        return flen(rbus[0].o_fine_ts)
-    else:
-        return 0
diff --git a/soc/runtime/Makefile b/soc/runtime/Makefile
index a5e0b51e8..0fcf8c7b3 100644
--- a/soc/runtime/Makefile
+++ b/soc/runtime/Makefile
@@ -54,7 +54,7 @@ ksupport.elf: $(OBJECTS_KSUPPORT)
 ksupport_data.o: ksupport.bin
 	$(LD) -r -b binary -o $@ $<
 
-service_table.h: $(SERVICE_TABLE_INPUT)
+service_table.h: $(SERVICE_TABLE_INPUT) gen_service_table.py
 	@echo " GEN     " $@ && ./gen_service_table.py $(SERVICE_TABLE_INPUT) > $@
 
 services.c: service_table.h
diff --git a/soc/runtime/gen_service_table.py b/soc/runtime/gen_service_table.py
index 5ef506abd..1e1e0aee1 100755
--- a/soc/runtime/gen_service_table.py
+++ b/soc/runtime/gen_service_table.py
@@ -6,11 +6,11 @@ import sys
 services = [
     ("syscalls", [
         ("rpc", "comm_rpc"),
-        ("rtio_oe", "rtio_oe"),
-        ("rtio_set", "rtio_set"),
+        ("rtio_set_o", "rtio_set_o"),
+        ("rtio_set_oe", "rtio_set_oe"),
+        ("rtio_set_sensitivity", "rtio_set_sensitivity"),
         ("rtio_get_counter", "rtio_get_counter"),
         ("rtio_get", "rtio_get"),
-        ("rtio_pileup_count", "rtio_pileup_count"),
         ("dds_phase_clear_en", "dds_phase_clear_en"),
         ("dds_program", "dds_program"),
     ]),
diff --git a/soc/runtime/rtio.c b/soc/runtime/rtio.c
index 441f062ee..9f156b5cf 100644
--- a/soc/runtime/rtio.c
+++ b/soc/runtime/rtio.c
@@ -18,19 +18,10 @@ void rtio_init(void)
     rtio_reset_write(0);
 }
 
-void rtio_oe(int channel, int oe)
-{
-    rtio_chan_sel_write(channel);
-    rtio_oe_write(oe);
-}
-
-void rtio_set(long long int timestamp, int channel, int value)
+static void write_and_process_status(long long int timestamp, int channel)
 {
     int status;
 
-    rtio_chan_sel_write(channel);
-    rtio_o_timestamp_write(timestamp);
-    rtio_o_value_write(value);
     rtio_o_we_write(1);
     status = rtio_o_status_read();
     if(status) {
@@ -49,6 +40,33 @@ void rtio_set(long long int timestamp, int channel, int value)
     }
 }
 
+void rtio_set_o(long long int timestamp, int channel, int value)
+{
+    rtio_chan_sel_write(channel);
+    rtio_o_timestamp_write(timestamp);
+    rtio_o_address_write(0);
+    rtio_o_data_write(value);
+    write_and_process_status(timestamp, channel);
+}
+
+void rtio_set_oe(long long int timestamp, int channel, int oe)
+{
+    rtio_chan_sel_write(channel);
+    rtio_o_timestamp_write(timestamp);
+    rtio_o_address_write(1);
+    rtio_o_data_write(oe);
+    write_and_process_status(timestamp, channel);
+}
+
+void rtio_set_sensitivity(long long int timestamp, int channel, int sensitivity)
+{
+    rtio_chan_sel_write(channel);
+    rtio_o_timestamp_write(timestamp);
+    rtio_o_address_write(2);
+    rtio_o_data_write(sensitivity);
+    write_and_process_status(timestamp, channel);
+}
+
 long long int rtio_get_counter(void)
 {
     rtio_counter_update_write(1);
@@ -81,16 +99,6 @@ long long int rtio_get(int channel, long long int time_limit)
     return r;
 }
 
-int rtio_pileup_count(int channel)
-{
-    int r;
-
-    rtio_chan_sel_write(channel);
-    r = rtio_i_pileup_count_read();
-    rtio_i_pileup_reset_write(1);
-    return r;
-}
-
 void rtio_fud_sync(void)
 {
     while(rtio_get_counter() < previous_fud_end_time);
@@ -102,14 +110,15 @@ void rtio_fud(long long int fud_time)
     int status;
 
     rtio_chan_sel_write(RTIO_FUD_CHANNEL);
+    rtio_o_address_write(0);
     fud_end_time = fud_time + 3*8;
     previous_fud_end_time = fud_end_time;
 
     rtio_o_timestamp_write(fud_time);
-    rtio_o_value_write(1);
+    rtio_o_data_write(1);
     rtio_o_we_write(1);
     rtio_o_timestamp_write(fud_end_time);
-    rtio_o_value_write(0);
+    rtio_o_data_write(0);
     rtio_o_we_write(1);
     status = rtio_o_status_read();
     if(status) {
diff --git a/soc/runtime/rtio.h b/soc/runtime/rtio.h
index 072236048..c1182e80b 100644
--- a/soc/runtime/rtio.h
+++ b/soc/runtime/rtio.h
@@ -2,11 +2,11 @@
 #define __RTIO_H
 
 void rtio_init(void);
-void rtio_oe(int channel, int oe);
-void rtio_set(long long int timestamp, int channel, int value);
+void rtio_set_o(long long int timestamp, int channel, int value);
+void rtio_set_oe(long long int timestamp, int channel, int oe);
+void rtio_set_sensitivity(long long int timestamp, int channel, int sensitivity);
 long long int rtio_get_counter(void);
 long long int rtio_get(int channel, long long int time_limit);
-int rtio_pileup_count(int channel);
 
 void rtio_fud_sync(void);
 void rtio_fud(long long int fud_time);
diff --git a/soc/runtime/test_mode.c b/soc/runtime/test_mode.c
index 82bbaff08..15a81d830 100644
--- a/soc/runtime/test_mode.c
+++ b/soc/runtime/test_mode.c
@@ -82,8 +82,8 @@ static void ttlout(char *n, char *value)
     }
 
     rtio_init();
-    rtio_oe(n2, 1);
-    rtio_set(rtio_get_counter() + 8000, n2, value2);
+    rtio_set_oe(rtio_get_counter() + 8000, n2, 1);
+    rtio_set_o(rtio_get_counter() + 8000, n2, value2);
 }
 
 static void ddssel(char *n)
diff --git a/soc/targets/artiq_kc705.py b/soc/targets/artiq_kc705.py
index 319b33c66..40ceed001 100644
--- a/soc/targets/artiq_kc705.py
+++ b/soc/targets/artiq_kc705.py
@@ -9,6 +9,7 @@ from misoclib.soc import mem_decoder
 from targets.kc705 import MiniSoC
 
 from artiq.gateware import amp, rtio, ad9858, nist_qc1
+from artiq.gateware.rtio.phy import ttl_simple
 
 
 class _RTIOCRG(Module, AutoCSR):
@@ -50,24 +51,36 @@ class _Peripherals(MiniSoC):
             platform.request("user_led", 0),
             platform.request("user_led", 1)))
 
-        fud = Signal()
         self.comb += [
             platform.request("ttl_l_tx_en").eq(1),
             platform.request("ttl_h_tx_en").eq(1)
         ]
-        rtio_ins = [platform.request("pmt") for i in range(2)]
-        rtio_outs = [platform.request("ttl", i) for i in range(16)]
-        rtio_outs.append(platform.request("user_led", 2))
-        self.add_constant("RTIO_FUD_CHANNEL", len(rtio_ins) + len(rtio_outs))
-        rtio_outs.append(fud)
 
+        # RTIO channels
+        rtio_channels = []
+        for i in range(2):
+            phy = ttl_simple.Inout(platform.request("pmt", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink, ififo_depth=512))
+        for i in range(16):
+            phy = ttl_simple.Output(platform.request("ttl", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        phy = ttl_simple.Output(platform.request("user_led", 2))
+        self.submodules += phy
+        rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        fud = Signal()
+        self.add_constant("RTIO_FUD_CHANNEL", len(rtio_channels))
+        phy = ttl_simple.Output(fud)
+        self.submodules += phy
+        rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        # RTIO core
         self.submodules.rtiocrg = _RTIOCRG(platform, self.crg.pll_sys)
-        self.submodules.rtiophy = rtio.phy.SimplePHY(
-            rtio_ins + rtio_outs,
-            output_only_pads=set(rtio_outs))
-        self.submodules.rtio = rtio.RTIO(self.rtiophy,
-                                         clk_freq=125000000,
-                                         ififo_depth=512)
+        self.submodules.rtio = rtio.RTIO(rtio_channels,
+                                         clk_freq=125000000)
 
         dds_pads = platform.request("dds")
         self.submodules.dds = ad9858.AD9858(dds_pads)
@@ -85,7 +98,7 @@ class UP(_Peripherals):
     def __init__(self, *args, **kwargs):
         _Peripherals.__init__(self, *args, **kwargs)
 
-        rtio_csrs = self.rtio.get_csrs()
+        rtio_csrs = self.rtio.get_csrs() + self.rtio.get_kernel_csrs()
         self.submodules.rtiowb = wbgen.Bank(rtio_csrs)
         self.add_wb_slave(mem_decoder(self.mem_map["rtio"]), self.rtiowb.bus)
         self.add_csr_region("rtio", self.mem_map["rtio"] + 0x80000000, 32, rtio_csrs)
@@ -93,6 +106,7 @@ class UP(_Peripherals):
         self.add_wb_slave(mem_decoder(self.mem_map["dds"]), self.dds.bus)
         self.add_memory_region("dds", self.mem_map["dds"] + 0x80000000, 64*4)
 
+
 class AMP(_Peripherals):
     csr_map = {
         "kernel_cpu": 14
@@ -112,7 +126,7 @@ class AMP(_Peripherals):
         self.add_wb_slave(mem_decoder(self.mem_map["mailbox"]), self.mailbox.i1)
         self.kernel_cpu.add_wb_slave(mem_decoder(self.mem_map["mailbox"]), self.mailbox.i2)
 
-        rtio_csrs = self.rtio.get_csrs()
+        rtio_csrs = self.rtio.get_kernel_csrs()
         self.submodules.rtiowb = wbgen.Bank(rtio_csrs)
         self.kernel_cpu.add_wb_slave(mem_decoder(self.mem_map["rtio"]), self.rtiowb.bus)
         self.add_csr_region("rtio", self.mem_map["rtio"] + 0x80000000, 32, rtio_csrs)
@@ -120,4 +134,5 @@ class AMP(_Peripherals):
         self.kernel_cpu.add_wb_slave(mem_decoder(self.mem_map["dds"]), self.dds.bus)
         self.add_memory_region("dds", self.mem_map["dds"] + 0x80000000, 64*4)
 
+
 default_subtarget = AMP
diff --git a/soc/targets/artiq_pipistrello.py b/soc/targets/artiq_pipistrello.py
index 5fdf507a1..c9b7415c6 100644
--- a/soc/targets/artiq_pipistrello.py
+++ b/soc/targets/artiq_pipistrello.py
@@ -7,6 +7,7 @@ from misoclib.soc import mem_decoder
 from targets.pipistrello import BaseSoC
 
 from artiq.gateware import amp, rtio, ad9858, nist_qc1
+from artiq.gateware.rtio.phy import ttl_simple
 
 
 class _RTIOCRG(Module, AutoCSR):
@@ -66,26 +67,46 @@ class _Peripherals(BaseSoC):
             platform.request("user_led", 1),
         ))
 
-        fud = Signal()
         self.comb += [
             platform.request("ttl_l_tx_en").eq(1),
             platform.request("ttl_h_tx_en").eq(1)
         ]
-        rtio_ins = [platform.request("pmt", i) for i in range(2)]
-        rtio_ins += [platform.request("xtrig", 0)]
-        rtio_outs = [platform.request("ttl", i) for i in range(16)]
-        rtio_outs += [platform.request("ext_led", 0)]
-        rtio_outs += [platform.request("user_led", i) for i in range(2, 5)]
-        self.add_constant("RTIO_FUD_CHANNEL", len(rtio_ins) + len(rtio_outs))
-        rtio_outs.append(fud)
 
+        # RTIO channels
+        rtio_channels = []
+        for i in range(2):
+            phy = ttl_simple.Inout(platform.request("pmt", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink, ififo_depth=512))
+
+        phy = ttl_simple.Inout(platform.request("xtrig", 0))
+        self.submodules += phy
+        rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        for i in range(16):
+            phy = ttl_simple.Output(platform.request("ttl", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        phy = ttl_simple.Output(platform.request("ext_led", 0))
+        self.submodules += phy
+        rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        for i in range(2, 5):
+            phy = ttl_simple.Output(platform.request("user_led", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        fud = Signal()
+        self.add_constant("RTIO_FUD_CHANNEL", len(rtio_channels))
+        phy = ttl_simple.Output(fud)
+        self.submodules += phy
+        rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        # RTIO core
         self.submodules.rtiocrg = _RTIOCRG(platform)
-        self.submodules.rtiophy = rtio.phy.SimplePHY(
-            rtio_ins + rtio_outs,
-            output_only_pads=set(rtio_outs))
-        self.submodules.rtio = rtio.RTIO(self.rtiophy,
-                                         clk_freq=125000000,
-                                         ififo_depth=512)
+        self.submodules.rtio = rtio.RTIO(rtio_channels,
+                                         clk_freq=125000000)
 
         dds_pads = platform.request("dds")
         self.submodules.dds = ad9858.AD9858(dds_pads)
diff --git a/soc/targets/artiq_ppro.py b/soc/targets/artiq_ppro.py
index 867a84796..8052effac 100644
--- a/soc/targets/artiq_ppro.py
+++ b/soc/targets/artiq_ppro.py
@@ -9,6 +9,7 @@ from misoclib.mem.sdram.core.minicon import MiniconSettings
 from targets.ppro import BaseSoC
 
 from artiq.gateware import rtio, ad9858, nist_qc1
+from artiq.gateware.rtio.phy import ttl_simple
 
 
 class _TestGen(Module):
@@ -78,29 +79,39 @@ class UP(BaseSoC):
             platform.request("user_led", 0),
             platform.request("ext_led", 0)))
 
-        fud = Signal()
         self.comb += [
             platform.request("ttl_l_tx_en").eq(1),
             platform.request("ttl_h_tx_en").eq(1)
         ]
-        rtio_ins = [platform.request("pmt") for i in range(2)]
-        rtio_outs = [platform.request("ttl", i) for i in range(5)]
-        self.add_constant("RTIO_FUD_CHANNEL", len(rtio_ins) + len(rtio_outs))
-        rtio_outs.append(fud)
 
+        # RTIO channels
+        rtio_channels = []
+        for i in range(2):
+            phy = ttl_simple.Inout(platform.request("pmt", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink))
+        for i in range(5):
+            phy = ttl_simple.Output(platform.request("ttl", i))
+            self.submodules += phy
+            rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        fud = Signal()
+        self.add_constant("RTIO_FUD_CHANNEL", len(rtio_channels))
+        phy = ttl_simple.Output(fud)
+        self.submodules += phy
+        rtio_channels.append(rtio.Channel(phy.rtlink))
+
+        # RTIO core
         self.submodules.rtiocrg = _RTIOMiniCRG(platform)
-        self.submodules.rtiophy = rtio.phy.SimplePHY(
-            rtio_ins + rtio_outs,
-            output_only_pads=set(rtio_outs))
-        self.submodules.rtio = rtio.RTIO(self.rtiophy,
+        self.submodules.rtio = rtio.RTIO(rtio_channels,
                                          clk_freq=125000000,
-                                         counter_width=32,
-                                         ififo_depth=512)
+                                         counter_width=32)
 
-        rtio_csrs = self.rtio.get_csrs()
+        rtio_csrs = self.rtio.get_csrs() + self.rtio.get_kernel_csrs()
         self.submodules.rtiowb = wbgen.Bank(rtio_csrs)
         self.add_wb_slave(mem_decoder(self.mem_map["rtio"]), self.rtiowb.bus)
-        self.add_csr_region("rtio", self.mem_map["rtio"] + 0x80000000, 32, rtio_csrs)
+        self.add_csr_region("rtio", self.mem_map["rtio"] + 0x80000000,
+                            32, rtio_csrs)
 
         if with_test_gen:
             self.submodules.test_gen = _TestGen(platform.request("ttl", 8))