diff --git a/artiq/coredevice/dds.py b/artiq/coredevice/dds.py
index f9af72eb4..6de0fbb0a 100644
--- a/artiq/coredevice/dds.py
+++ b/artiq/coredevice/dds.py
@@ -1,10 +1,10 @@
 from artiq.language.core import *
 from artiq.language.db import *
 from artiq.language.units import *
-from artiq.coredevice import ttl
 
 
 PHASE_MODE_DEFAULT = -1
+# keep in sync with dds.h
 PHASE_MODE_CONTINUOUS = 0
 PHASE_MODE_ABSOLUTE = 1
 PHASE_MODE_TRACKING = 2
@@ -13,34 +13,24 @@ PHASE_MODE_TRACKING = 2
 class DDS(AutoDB):
     """Core device Direct Digital Synthesis (DDS) driver.
 
-    Controls DDS devices managed directly by the core device's runtime. It also
-    uses a RTIO TTL channel (through :class:`artiq.coredevice.ttl.TTLOut`) to
-    control a RF switch that gates the output of the DDS device.
+    Controls DDS devices managed directly by the core device's runtime.
 
     :param dds_sysclk: DDS system frequency, used for computing the frequency
         tuning words.
-    :param reg_channel: channel number of the DDS device to control.
-    :param rtio_switch: RTIO channel number of the RF switch associated with
-        the DDS device.
-
+    :param channel: channel number of the DDS device to control.
     """
     class DBKeys:
         core = Device()
-        dds_sysclk = Parameter(1*GHz)
-        reg_channel = Argument()
-        rtio_switch = Argument()
+        dds_sysclk = Argument(1*GHz)
+        channel = Argument()
 
     def build(self):
-        self.previous_on = False
-        self.previous_frequency = 0*MHz
-        self.set_phase_mode(PHASE_MODE_CONTINUOUS)
-        self.sw = ttl.TTLOut(core=self.core, channel=self.rtio_switch)
+        self.phase_mode = PHASE_MODE_CONTINUOUS
 
     @portable
     def frequency_to_ftw(self, frequency):
         """Returns the frequency tuning word corresponding to the given
         frequency.
-
         """
         return round(2**32*frequency/self.dds_sysclk)
 
@@ -48,10 +38,14 @@ class DDS(AutoDB):
     def ftw_to_frequency(self, ftw):
         """Returns the frequency corresponding to the given frequency tuning
         word.
-
         """
         return ftw*self.dds_sysclk/2**32
 
+    @kernel
+    def init(self):
+        """Resets and initializes the DDS."""
+        syscall("dds_init", time_to_cycles(now()), self.channel)
+
     @kernel
     def set_phase_mode(self, phase_mode):
         """Sets the phase mode of the DDS channel. Supported phase modes are:
@@ -69,74 +63,21 @@ class DDS(AutoDB):
           accumulator is set to the value it would have if the DDS had been
           running at the specified frequency since the start of the
           experiment.
-
         """
         self.phase_mode = phase_mode
-        syscall("dds_phase_clear_en", self.reg_channel,
-                self.phase_mode != PHASE_MODE_CONTINUOUS)
 
     @kernel
-    def on(self, frequency, phase_mode=PHASE_MODE_DEFAULT, phase_offset=0):
-        """Sets the DDS channel to the specified frequency and turns it on.
-
-        If the DDS channel was already on, a real-time frequency and phase
-        update is performed.
+    def set(self, frequency, phase_mode=PHASE_MODE_DEFAULT, phase_offset=0):
+        """Sets the DDS channel to the specified frequency and phase.
 
         :param frequency: frequency to generate.
         :param phase_mode: if specified, overrides the default phase mode set
             by ``set_phase_mode`` for this call.
         :param phase_offset: adds an offset, in turns, to the phase.
-
         """
-        if phase_mode != PHASE_MODE_DEFAULT:
-            old_phase_mode = self.phase_mode
-            self.set_phase_mode(phase_mode)
+        if phase_mode == PHASE_MODE_DEFAULT:
+            phase_mode = self.phase_mode
 
-        if self.previous_frequency != frequency:
-            merge = self.sw.o_previous_timestamp == time_to_cycles(now())
-            if not merge:
-                self.sw.sync()
-            # Channel is already on:
-            # Precise timing of frequency change is required.
-            # Channel is off:
-            # Use soft timing on FUD to prevent conflicts when reprogramming
-            # several channels that need to be turned on at the same time.
-            rt_fud = merge or self.previous_on
-            if self.phase_mode != PHASE_MODE_CONTINUOUS:
-                sysclk_per_microcycle = int(self.dds_sysclk*
-                                            self.core.ref_period)
-            else:
-                sysclk_per_microcycle = 0
-            syscall("dds_program", time_to_cycles(now()), self.reg_channel,
-               self.frequency_to_ftw(frequency), int(phase_offset*2**14),
-               sysclk_per_microcycle,
-               rt_fud, self.phase_mode == PHASE_MODE_TRACKING)
-            self.previous_frequency = frequency
-        self.sw.on()
-        self.previous_on = True
-
-        if phase_mode != PHASE_MODE_DEFAULT:
-            self.set_phase_mode(old_phase_mode)
-
-    @kernel
-    def off(self):
-        """Turns the DDS channel off.
-
-        """
-        self.sw.off()
-        self.previous_on = False
-
-    @kernel
-    def pulse(self, frequency, duration,
-              phase_mode=PHASE_MODE_DEFAULT, phase_offset=0):
-        """Pulses the DDS channel for the specified duration at the specified
-        frequency.
-
-        See ``on`` for a description of the parameters.
-
-        Equivalent to a ``on``, ``delay``, ``off`` sequence.
-
-        """
-        self.on(frequency, phase_mode, phase_offset)
-        delay(duration)
-        self.off()
+        syscall("dds_set", time_to_cycles(now()), self.channel,
+           self.frequency_to_ftw(frequency), round(phase_offset*2**14),
+           self.phase_mode)
diff --git a/artiq/coredevice/runtime.py b/artiq/coredevice/runtime.py
index 3d84decc8..fb6bf7f1d 100644
--- a/artiq/coredevice/runtime.py
+++ b/artiq/coredevice/runtime.py
@@ -21,8 +21,8 @@ _syscalls = {
     "rtio_set_sensitivity": "Iii:n",
     "rtio_get_counter": "n:I",
     "rtio_get": "iI:I",
-    "dds_phase_clear_en": "ib:n",
-    "dds_program": "Iiiiibb:n",
+    "dds_init": "Ii:n",
+    "dds_set": "Iiiii:n",
 }
 
 
diff --git a/artiq/gateware/ad9858.py b/artiq/gateware/ad9858.py
index 938d13d73..26c790672 100644
--- a/artiq/gateware/ad9858.py
+++ b/artiq/gateware/ad9858.py
@@ -36,8 +36,8 @@ class AD9858(Module):
 
     Read timing:
     Address is set one cycle before assertion of rd_n.
-    rd_n is asserted for 3 cycles.
-    Data is sampled 2 cycles into the assertion of rd_n.
+    rd_n is asserted for read_wait_cycles, data is sampled at the end.
+    rd_n is deasserted and data bus is not driven again before hiz_wait_cycles.
 
     Design:
     All IO pads are registered.
@@ -48,7 +48,7 @@ class AD9858(Module):
     Round-trip addr A setup (> RX, RD, D to Z), RD prop, D valid (< D
     valid), D prop is ~15 + 10 + 20 + 10 = 55ns
     """
-    def __init__(self, pads, drive_fud=False,
+    def __init__(self, pads,
                  read_wait_cycles=10, hiz_wait_cycles=3,
                  bus=None):
         if bus is None:
@@ -84,9 +84,8 @@ class AD9858(Module):
                 bus.dat_r.eq(dr)
             )
 
-        if drive_fud:
-            fud = Signal()
-            self.sync += pads.fud_n.eq(~fud)
+        fud = Signal()
+        self.sync += pads.fud_n.eq(~fud)
 
         pads.wr_n.reset = 1
         pads.rd_n.reset = 1
@@ -106,7 +105,7 @@ class AD9858(Module):
                     If(bus.adr[0],
                         NextState("GPIO")
                     ).Else(
-                        NextState("FUD") if drive_fud else None
+                        NextState("FUD")
                     )
                 ).Else(
                     If(bus.we,
@@ -141,19 +140,18 @@ class AD9858(Module):
         )
         fsm.act("WAIT_HIZ",
             rx.eq(1),
-            # For some reason, AD9858 has a address hold time to RD inactive.
+            # For some reason, AD9858 has an address hold time to RD inactive.
             hold_address.eq(1),
             self.hiz_timer.wait.eq(1),
             If(self.hiz_timer.done, NextState("IDLE"))
         )
-        if drive_fud:
-            fsm.act("FUD",
-                # 4ns FUD setup to SYNCLK
-                # 0ns FUD hold to SYNCLK
-                fud.eq(1),
-                bus.ack.eq(1),
-                NextState("IDLE")
-            )
+        fsm.act("FUD",
+            # 4ns FUD setup to SYNCLK
+            # 0ns FUD hold to SYNCLK
+            fud.eq(1),
+            bus.ack.eq(1),
+            NextState("IDLE")
+        )
         fsm.act("GPIO",
             bus.ack.eq(1),
             bus_r_gpio.eq(1),
diff --git a/artiq/gateware/rtio/core.py b/artiq/gateware/rtio/core.py
index 946ded34d..cfa7ea01e 100644
--- a/artiq/gateware/rtio/core.py
+++ b/artiq/gateware/rtio/core.py
@@ -291,6 +291,10 @@ class RTIO(Module):
         fine_ts_width = max(rtlink.get_fine_ts_width(c.interface)
                             for c in channels)
 
+        self.data_width = data_width
+        self.address_width = address_width
+        self.fine_ts_width = fine_ts_width
+
         # CSRs
         self.kcsrs = _KernelCSRs(bits_for(len(channels)-1),
                                  data_width, address_width,
diff --git a/artiq/gateware/rtio/phy/wishbone.py b/artiq/gateware/rtio/phy/wishbone.py
index 773dcb9bf..b1184ca43 100644
--- a/artiq/gateware/rtio/phy/wishbone.py
+++ b/artiq/gateware/rtio/phy/wishbone.py
@@ -1,10 +1,14 @@
 from migen.fhdl.std import *
+from migen.bus import wishbone
 
 from artiq.gateware.rtio import rtlink
 
 
 class RT2WB(Module):
-    def __init__(self, wb, address_width):
+    def __init__(self, address_width, wb=None):
+        if wb is None:
+            wb = wishbone.Interface()
+        self.wb = wb
         self.rtlink = rtlink.Interface(
             rtlink.OInterface(
                 flen(wb.dat_w),
@@ -18,7 +22,7 @@ class RT2WB(Module):
         # # #
 
         active = Signal()
-        self.sync.rio_phy += [
+        self.sync.rio += [
             If(self.rtlink.o.stb,
                 active.eq(1),
                 wb.adr.eq(self.rtlink.o.address[:address_width]),
@@ -35,6 +39,6 @@ class RT2WB(Module):
             wb.cyc.eq(active),
             wb.stb.eq(active),
 
-            self.i.stb.eq(wb.ack & ~wb.we),
-            self.i.data.eq(wb.dat_r)
+            self.rtlink.i.stb.eq(wb.ack & ~wb.we),
+            self.rtlink.i.data.eq(wb.dat_r)
         ]
diff --git a/examples/master/ddb.pyon b/examples/master/ddb.pyon
index a33fc8bec..618341244 100644
--- a/examples/master/ddb.pyon
+++ b/examples/master/ddb.pyon
@@ -54,19 +54,19 @@
         "type": "local",
         "module": "artiq.coredevice.dds",
         "class": "DDS",
-        "arguments": {"reg_channel": 0, "rtio_switch": 5}
+        "arguments": {"channel": 0}
     },
     "dds1": {
         "type": "local",
         "module": "artiq.coredevice.dds",
         "class": "DDS",
-        "arguments": {"reg_channel": 1, "rtio_switch": 6}
+        "arguments": {"channel": 1}
     },
     "dds2": {
         "type": "local",
         "module": "artiq.coredevice.dds",
         "class": "DDS",
-        "arguments": {"reg_channel": 2, "rtio_switch": 7}
+        "arguments": {"channel": 2}
     },
 
     "qc_q1_0": {
diff --git a/examples/master/repository/dds_test.py b/examples/master/repository/dds_test.py
index 7a18d6620..520faa5ce 100644
--- a/examples/master/repository/dds_test.py
+++ b/examples/master/repository/dds_test.py
@@ -9,10 +9,17 @@ class DDSTest(Experiment, AutoDB):
         dds0 = Device()
         dds1 = Device()
         dds2 = Device()
+        ttl0 = Device()
+        ttl1 = Device()
+        ttl2 = Device()
         led = Device()
 
     @kernel
     def run(self):
+        # with dds_batch:
+        #     self.dds1.set(120*MHz)
+        #     self.dds2.set(200*MHz)
+
         for i in range(10000):
             if i & 0x200:
                 self.led.on()
@@ -20,7 +27,8 @@ class DDSTest(Experiment, AutoDB):
                 self.led.off()
             with parallel:
                 with sequential:
-                    self.dds0.pulse(100*MHz + 4*i*kHz, 500*us)
-                    self.dds1.pulse(120*MHz, 500*us)
-                self.dds2.pulse(200*MHz, 100*us)
+                    self.dds0.set(100*MHz + 4*i*kHz)
+                    self.ttl0.pulse(500*us)
+                    self.ttl1.pulse(500*us)
+                self.ttl2.pulse(100*us)
         self.led.off()
diff --git a/soc/runtime/Makefile b/soc/runtime/Makefile
index a27e651f8..ba75870fc 100644
--- a/soc/runtime/Makefile
+++ b/soc/runtime/Makefile
@@ -34,7 +34,9 @@ ksupport.elf: $(OBJECTS_KSUPPORT)
 		-T ksupport.ld \
 		-N -o $@ \
 		$(MSCDIR)/software/libbase/crt0-$(CPU).o \
-		$^
+		$^ \
+		-L$(MSCDIR)/software/libcompiler-rt \
+		-lcompiler-rt
 	@chmod -x $@
 
 ksupport_data.o: ksupport.bin
diff --git a/soc/runtime/bridge.c b/soc/runtime/bridge.c
index 68da7fa64..d467a9a22 100644
--- a/soc/runtime/bridge.c
+++ b/soc/runtime/bridge.c
@@ -4,6 +4,26 @@
 #include "dds.h"
 #include "bridge.h"
 
+static void dds_write(int addr, int data)
+{
+    rtio_chan_sel_write(RTIO_DDS_CHANNEL);
+    rtio_o_address_write(addr);
+    rtio_o_data_write(data);
+    rtio_o_timestamp_write(rtio_get_counter() + 8000);
+    rtio_o_we_write(1);
+}
+
+static int dds_read(int addr)
+{
+    int r;
+
+    dds_write(addr | 128, 0);
+    while(rtio_i_status_read() & RTIO_I_STATUS_EMPTY);
+    r = rtio_i_data_read();
+    rtio_i_re_write(1);
+    return r;
+}
+
 static void send_ready(void)
 {
     struct msg_base msg;
@@ -41,16 +61,16 @@ void bridge_main(void)
                 struct msg_brg_dds_sel *msg;
 
                 msg = (struct msg_brg_dds_sel *)umsg;
-                DDS_WRITE(DDS_GPIO, msg->channel);
+                dds_write(DDS_GPIO, msg->channel);
                 mailbox_acknowledge();
                 break;
             }
             case MESSAGE_TYPE_BRG_DDS_RESET: {
                 unsigned int g;
 
-                g = DDS_READ(DDS_GPIO);
-                DDS_WRITE(DDS_GPIO, g | (1 << 7));
-                DDS_WRITE(DDS_GPIO, g);
+                g = dds_read(DDS_GPIO);
+                dds_write(DDS_GPIO, g | (1 << 7));
+                dds_write(DDS_GPIO, g);
 
                 mailbox_acknowledge();
                 break;
@@ -61,7 +81,7 @@ void bridge_main(void)
 
                 msg = (struct msg_brg_dds_read_request *)umsg;
                 rmsg.type = MESSAGE_TYPE_BRG_DDS_READ_REPLY;
-                rmsg.data = DDS_READ(msg->address);
+                rmsg.data = dds_read(msg->address);
                 mailbox_send_and_wait(&rmsg);
                 break;
             }
@@ -69,12 +89,12 @@ void bridge_main(void)
                 struct msg_brg_dds_write *msg;
 
                 msg = (struct msg_brg_dds_write *)umsg;
-                DDS_WRITE(msg->address, msg->data);
+                dds_write(msg->address, msg->data);
                 mailbox_acknowledge();
                 break;
             }
             case MESSAGE_TYPE_BRG_DDS_FUD:
-                rtio_fud(rtio_get_counter() + 8000);
+                dds_write(DDS_FUD, 0);
                 mailbox_acknowledge();
                 break;
             default:
diff --git a/soc/runtime/dds.c b/soc/runtime/dds.c
index 1d4bf5438..87aa8464c 100644
--- a/soc/runtime/dds.c
+++ b/soc/runtime/dds.c
@@ -4,64 +4,66 @@
 #include "rtio.h"
 #include "dds.h"
 
-void dds_init(void)
-{
-    int i;
+#define DURATION_WRITE 5
+#define DURATION_PROGRAM (8*DURATION_WRITE)
 
-    for(i=0;i<8;i++) {
-        DDS_WRITE(DDS_GPIO, i);
-        DDS_WRITE(DDS_GPIO, i | (1 << 7));
-        DDS_WRITE(DDS_GPIO, i);
-        dds_phase_clear_en(i, 0);
-    }
-}
+#define DDS_WRITE(addr, data) do { \
+        rtio_o_address_write(addr); \
+        rtio_o_data_write(data); \
+        rtio_o_timestamp_write(now); \
+        rtio_write_and_process_status(now, RTIO_DDS_CHANNEL); \
+        now += DURATION_WRITE; \
+    } while(0)
 
-void dds_phase_clear_en(int channel, int phase_clear_en)
+void dds_init(long long int timestamp, int channel)
 {
+    long long int now;
+
+    rtio_chan_sel_write(RTIO_DDS_CHANNEL);
+
+    now = timestamp - 7*DURATION_WRITE;
+
+    DDS_WRITE(DDS_GPIO, channel);
+    DDS_WRITE(DDS_GPIO, channel | (1 << 7));
+    DDS_WRITE(DDS_GPIO, channel);
+
     DDS_WRITE(0x00, 0x78);
     DDS_WRITE(0x01, 0x00);
-    DDS_WRITE(0x02, phase_clear_en ? 0x40 : 0x00);
+    DDS_WRITE(0x02, 0x00);
     DDS_WRITE(0x03, 0x00);
 }
 
-/*
- * DDS phase modes:
- * - continuous: Set sysclk_per_microcycle=0 to disable POW alteration.
- *               phase_tracking is ignored, set to 0.
- *               Disable phase accumulator clearing prior to programming.
- * - absolute:   Set sysclk_per_microcycle to its nominal value
- *               and phase_tracking=0.
- *               Enable phase accumulator clearing prior to programming.
- * - tracking:   Set sysclk_per_microcycle to its nominal value
- *               and phase_tracking=1.
-*                Enable phase accumulator clearing prior to programming.
- */
-void dds_program(long long int timestamp, int channel,
-    unsigned int ftw, unsigned int pow, unsigned int sysclk_per_microcycle,
-    int rt_fud, int phase_tracking)
+static void dds_set_one(long long int now, long long int timestamp, int channel,
+    unsigned int ftw, unsigned int pow, int phase_mode)
 {
-    long long int fud_time;
-    unsigned int phase_time_offset;
-
-    rtio_fud_sync();
     DDS_WRITE(DDS_GPIO, channel);
 
+    if(phase_mode == PHASE_MODE_CONTINUOUS)
+        /* Do not clear phase accumulator on FUD */
+        DDS_WRITE(0x02, 0x00);
+    else
+        /* Clear phase accumulator on FUD */
+        DDS_WRITE(0x02, 0x40);
+
     DDS_WRITE(DDS_FTW0, ftw & 0xff);
     DDS_WRITE(DDS_FTW1, (ftw >> 8) & 0xff);
     DDS_WRITE(DDS_FTW2, (ftw >> 16) & 0xff);
     DDS_WRITE(DDS_FTW3, (ftw >> 24) & 0xff);
 
-    phase_time_offset = phase_tracking ? timestamp : 0;
-    if(rt_fud)
-        fud_time = timestamp;
-    else {
-        fud_time = rtio_get_counter() + 8000;
-        /* POW is mod 2**14, so wraparound on negative values is OK */
-        phase_time_offset -= timestamp - fud_time;
-    }
-    pow += phase_time_offset*ftw*sysclk_per_microcycle >> 18;
+    if(phase_mode == PHASE_MODE_TRACKING)
+        /* We assume that the RTIO clock is DDS SYNCLK */
+        pow += (timestamp >> RTIO_FINE_TS_WIDTH)*ftw >> 18;
     DDS_WRITE(DDS_POW0, pow & 0xff);
     DDS_WRITE(DDS_POW1, (pow >> 8) & 0x3f);
-
-    rtio_fud(fud_time);
+}
+
+void dds_set(long long int timestamp, int channel,
+    unsigned int ftw, unsigned int pow, int phase_mode)
+{
+    long long int now;
+
+    rtio_chan_sel_write(RTIO_DDS_CHANNEL);
+    dds_set_one(timestamp - DURATION_PROGRAM, timestamp, channel, ftw, pow, phase_mode);
+    now = timestamp;
+    DDS_WRITE(DDS_FUD, 0);
 }
diff --git a/soc/runtime/dds.h b/soc/runtime/dds.h
index fdbd8103f..c38935277 100644
--- a/soc/runtime/dds.h
+++ b/soc/runtime/dds.h
@@ -4,24 +4,23 @@
 #include <hw/common.h>
 #include <generated/mem.h>
 
-#define DDS_READ(addr) \
-    MMPTR(DDS_BASE + (addr)*4)
-
-#define DDS_WRITE(addr, data) \
-    MMPTR(DDS_BASE + (addr)*4) = data
-
 #define DDS_FTW0  0x0a
 #define DDS_FTW1  0x0b
 #define DDS_FTW2  0x0c
 #define DDS_FTW3  0x0d
 #define DDS_POW0  0x0e
 #define DDS_POW1  0x0f
+#define DDS_FUD   0x40
 #define DDS_GPIO  0x41
 
-void dds_init(void);
-void dds_phase_clear_en(int channel, int phase_clear_en);
-void dds_program(long long int timestamp, int channel,
-    unsigned int ftw, unsigned int pow, unsigned int sysclk_per_microcycle,
-    int rt_fud, int phase_tracking);
+enum {
+    PHASE_MODE_CONTINUOUS = 0,
+    PHASE_MODE_ABSOLUTE = 1,
+    PHASE_MODE_TRACKING = 2
+};
+
+void dds_init(long long int timestamp, int channel);
+void dds_set(long long int timestamp, int channel,
+    unsigned int ftw, unsigned int pow, int phase_mode);
 
 #endif /* __DDS_H */
diff --git a/soc/runtime/gen_service_table.py b/soc/runtime/gen_service_table.py
index c4731be8f..a58236207 100755
--- a/soc/runtime/gen_service_table.py
+++ b/soc/runtime/gen_service_table.py
@@ -17,8 +17,8 @@ services = [
         ("rtio_set_sensitivity", "rtio_set_sensitivity"),
         ("rtio_get_counter", "rtio_get_counter"),
         ("rtio_get", "rtio_get"),
-        ("dds_phase_clear_en", "dds_phase_clear_en"),
-        ("dds_program", "dds_program"),
+        ("dds_init", "dds_init"),
+        ("dds_set", "dds_set"),
     ]),
     ("eh", [
         ("setjmp", "exception_setjmp"),
diff --git a/soc/runtime/ksupport.c b/soc/runtime/ksupport.c
index 964391341..a2eb090c6 100644
--- a/soc/runtime/ksupport.c
+++ b/soc/runtime/ksupport.c
@@ -78,7 +78,6 @@ long long int now_init(void)
     mailbox_acknowledge();
 
     if(now < 0) {
-        dds_init();
         rtio_init();
         now = 125000;
     }
diff --git a/soc/runtime/rtio.c b/soc/runtime/rtio.c
index e1a2e9610..8fdcd2759 100644
--- a/soc/runtime/rtio.c
+++ b/soc/runtime/rtio.c
@@ -3,42 +3,17 @@
 #include "exceptions.h"
 #include "rtio.h"
 
-#define RTIO_O_STATUS_FULL 1
-#define RTIO_O_STATUS_UNDERFLOW 2
-#define RTIO_O_STATUS_SEQUENCE_ERROR 4
-#define RTIO_I_STATUS_EMPTY 1
-#define RTIO_I_STATUS_OVERFLOW 2
-
-long long int previous_fud_end_time;
-
 void rtio_init(void)
 {
-    previous_fud_end_time = 0;
     rtio_reset_write(1);
     rtio_reset_write(0);
     rtio_reset_phy_write(0);
 }
 
-static void write_and_process_status(long long int timestamp, int channel)
+long long int rtio_get_counter(void)
 {
-    int status;
-
-    rtio_o_we_write(1);
-    status = rtio_o_status_read();
-    if(status) {
-        if(status & RTIO_O_STATUS_FULL)
-            while(rtio_o_status_read() & RTIO_O_STATUS_FULL);
-        if(status & RTIO_O_STATUS_UNDERFLOW) {
-            rtio_o_underflow_reset_write(1);
-            exception_raise_params(EID_RTIO_UNDERFLOW,
-                timestamp, channel, rtio_get_counter());
-        }
-        if(status & RTIO_O_STATUS_SEQUENCE_ERROR) {
-            rtio_o_sequence_error_reset_write(1);
-            exception_raise_params(EID_RTIO_SEQUENCE_ERROR,
-                timestamp, channel, 0);
-        }
-    }
+    rtio_counter_update_write(1);
+    return rtio_counter_read();
 }
 
 void rtio_set_o(long long int timestamp, int channel, int value)
@@ -47,7 +22,7 @@ void rtio_set_o(long long int timestamp, int channel, int value)
     rtio_o_timestamp_write(timestamp);
     rtio_o_address_write(0);
     rtio_o_data_write(value);
-    write_and_process_status(timestamp, channel);
+    rtio_write_and_process_status(timestamp, channel);
 }
 
 void rtio_set_oe(long long int timestamp, int channel, int oe)
@@ -56,7 +31,7 @@ void rtio_set_oe(long long int timestamp, int channel, int oe)
     rtio_o_timestamp_write(timestamp);
     rtio_o_address_write(1);
     rtio_o_data_write(oe);
-    write_and_process_status(timestamp, channel);
+    rtio_write_and_process_status(timestamp, channel);
 }
 
 void rtio_set_sensitivity(long long int timestamp, int channel, int sensitivity)
@@ -65,13 +40,7 @@ void rtio_set_sensitivity(long long int timestamp, int channel, int sensitivity)
     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);
-    return rtio_counter_read();
+    rtio_write_and_process_status(timestamp, channel);
 }
 
 long long int rtio_get(int channel, long long int time_limit)
@@ -99,39 +68,3 @@ long long int rtio_get(int channel, long long int time_limit)
     rtio_i_re_write(1);
     return r;
 }
-
-void rtio_fud_sync(void)
-{
-    while(rtio_get_counter() < previous_fud_end_time);
-}
-
-void rtio_fud(long long int fud_time)
-{
-    long long int fud_end_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_data_write(1);
-    rtio_o_we_write(1);
-    rtio_o_timestamp_write(fud_end_time);
-    rtio_o_data_write(0);
-    rtio_o_we_write(1);
-    status = rtio_o_status_read();
-    if(status) {
-        if(status & RTIO_O_STATUS_UNDERFLOW) {
-            rtio_o_underflow_reset_write(1);
-            exception_raise_params(EID_RTIO_UNDERFLOW,
-                fud_time, RTIO_FUD_CHANNEL, rtio_get_counter());
-        }
-        if(status & RTIO_O_STATUS_SEQUENCE_ERROR) {
-            rtio_o_sequence_error_reset_write(1);
-            exception_raise_params(EID_RTIO_SEQUENCE_ERROR,
-                fud_time, RTIO_FUD_CHANNEL, 0);
-        }
-    }
-}
diff --git a/soc/runtime/rtio.h b/soc/runtime/rtio.h
index c1182e80b..6c30cec12 100644
--- a/soc/runtime/rtio.h
+++ b/soc/runtime/rtio.h
@@ -1,14 +1,43 @@
 #ifndef __RTIO_H
 #define __RTIO_H
 
+#include <generated/csr.h>
+#include "exceptions.h"
+
+#define RTIO_O_STATUS_FULL 1
+#define RTIO_O_STATUS_UNDERFLOW 2
+#define RTIO_O_STATUS_SEQUENCE_ERROR 4
+#define RTIO_I_STATUS_EMPTY 1
+#define RTIO_I_STATUS_OVERFLOW 2
+
 void rtio_init(void);
+long long int rtio_get_counter(void);
+
+static inline void rtio_write_and_process_status(long long int timestamp, int channel)
+{
+    int status;
+
+    rtio_o_we_write(1);
+    status = rtio_o_status_read();
+    if(status) {
+        if(status & RTIO_O_STATUS_FULL)
+            while(rtio_o_status_read() & RTIO_O_STATUS_FULL);
+        if(status & RTIO_O_STATUS_UNDERFLOW) {
+            rtio_o_underflow_reset_write(1);
+            exception_raise_params(EID_RTIO_UNDERFLOW,
+                timestamp, channel, rtio_get_counter());
+        }
+        if(status & RTIO_O_STATUS_SEQUENCE_ERROR) {
+            rtio_o_sequence_error_reset_write(1);
+            exception_raise_params(EID_RTIO_SEQUENCE_ERROR,
+                timestamp, channel, 0);
+        }
+    }
+}
+
 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);
 
-void rtio_fud_sync(void);
-void rtio_fud(long long int fud_time);
-
 #endif /* __RTIO_H */
diff --git a/soc/targets/artiq_kc705.py b/soc/targets/artiq_kc705.py
index 94fe5fc80..b42f445c7 100644
--- a/soc/targets/artiq_kc705.py
+++ b/soc/targets/artiq_kc705.py
@@ -11,6 +11,7 @@ from targets.kc705 import MiniSoC
 from artiq.gateware.soc import AMPSoC
 from artiq.gateware import rtio, ad9858, nist_qc1
 from artiq.gateware.rtio.phy import ttl_simple
+from artiq.gateware.rtio.phy.wishbone import RT2WB
 
 
 class _RTIOCRG(Module, AutoCSR):
@@ -40,7 +41,6 @@ class NIST_QC1(MiniSoC, AMPSoC):
     csr_map.update(MiniSoC.csr_map)
     mem_map = {
         "rtio":     0x20000000, # (shadow @0xa0000000)
-        "dds":      0x50000000, # (shadow @0xd0000000)
         "mailbox":  0x70000000  # (shadow @0xf0000000)
     }
     mem_map.update(MiniSoC.mem_map)
@@ -75,20 +75,20 @@ class NIST_QC1(MiniSoC, AMPSoC):
         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.add_constant("RTIO_DDS_CHANNEL", len(rtio_channels))
+        self.submodules.dds = RenameClockDomains(
+            ad9858.AD9858(platform.request("dds")),
+            "rio")
+        phy = RT2WB(7, self.dds.bus)
         self.submodules += phy
-        rtio_channels.append(rtio.Channel(phy.rtlink))
+        rtio_channels.append(rtio.Channel(phy.rtlink, ififo_depth=4))
 
         # RTIO core
         self.submodules.rtiocrg = _RTIOCRG(platform, self.crg.pll_sys)
         self.submodules.rtio = rtio.RTIO(rtio_channels,
                                          clk_freq=125000000)
+        self.add_constant("RTIO_FINE_TS_WIDTH", self.rtio.fine_ts_width)
 
-        dds_pads = platform.request("dds")
-        self.submodules.dds = ad9858.AD9858(dds_pads)
-        self.comb += dds_pads.fud_n.eq(~fud)
 
         if isinstance(platform.toolchain, XilinxVivadoToolchain):
             platform.add_platform_command("""
@@ -106,9 +106,5 @@ set_false_path -from [get_clocks rio_clk] -to [get_clocks rsys_clk]
         self.add_csr_region("rtio", self.mem_map["rtio"] | 0x80000000, 32,
                             rtio_csrs)
 
-        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 = NIST_QC1
diff --git a/soc/targets/artiq_pipistrello.py b/soc/targets/artiq_pipistrello.py
index fac40e2f6..29259d529 100644
--- a/soc/targets/artiq_pipistrello.py
+++ b/soc/targets/artiq_pipistrello.py
@@ -9,6 +9,7 @@ from targets.pipistrello import BaseSoC
 from artiq.gateware.soc import AMPSoC
 from artiq.gateware import rtio, ad9858, nist_qc1
 from artiq.gateware.rtio.phy import ttl_simple
+from artiq.gateware.rtio.phy.wishbone import RT2WB
 
 
 class _RTIOCRG(Module, AutoCSR):
@@ -61,7 +62,6 @@ class NIST_QC1(BaseSoC, AMPSoC):
     csr_map.update(BaseSoC.csr_map)
     mem_map = {
         "rtio":     0x20000000,  # (shadow @0xa0000000)
-        "dds":      0x50000000,  # (shadow @0xd0000000)
         "mailbox":  0x70000000   # (shadow @0xf0000000)
     }
     mem_map.update(BaseSoC.mem_map)
@@ -110,20 +110,19 @@ trce -v 12 -fastpaths -tsi {build_name}.tsi -o {build_name}.twr {build_name}.ncd
             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.add_constant("RTIO_DDS_CHANNEL", len(rtio_channels))
+        self.submodules.dds = RenameClockDomains(
+            ad9858.AD9858(platform.request("dds")),
+            "rio")
+        phy = RT2WB(7, self.dds.bus)
         self.submodules += phy
-        rtio_channels.append(rtio.Channel(phy.rtlink))
+        rtio_channels.append(rtio.Channel(phy.rtlink, ififo_depth=4))
 
         # RTIO core
         self.submodules.rtiocrg = _RTIOCRG(platform)
         self.submodules.rtio = rtio.RTIO(rtio_channels,
                                          clk_freq=125000000)
-
-        dds_pads = platform.request("dds")
-        self.submodules.dds = ad9858.AD9858(dds_pads)
-        self.comb += dds_pads.fud_n.eq(~fud)
+        self.add_constant("RTIO_FINE_TS_WIDTH", self.rtio.fine_ts_width)
 
         # CPU connections
         rtio_csrs = self.rtio.get_csrs()
@@ -133,9 +132,5 @@ trce -v 12 -fastpaths -tsi {build_name}.tsi -o {build_name}.twr {build_name}.ncd
         self.add_csr_region("rtio", self.mem_map["rtio"] | 0x80000000, 32,
                             rtio_csrs)
 
-        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 = NIST_QC1