diff --git a/runtime/runtime/src/Quidditch/dispatch/dispatch.c b/runtime/runtime/src/Quidditch/dispatch/dispatch.c
index 37b5361..dea494c 100644
--- a/runtime/runtime/src/Quidditch/dispatch/dispatch.c
+++ b/runtime/runtime/src/Quidditch/dispatch/dispatch.c
@@ -1,17 +1,19 @@
 
 #include "dispatch.h"
 
+#include <iree/base/alignment.h>
+
 #include <assert.h>
 #include <cluster_interrupt_decls.h>
 #include <encoding.h>
+#include <perf_decls.h>
 #include <riscv_decls.h>
+#include <ssr_decls.h>
 #include <stdatomic.h>
 #include <stdbool.h>
 #include <stdint.h>
 #include <team_decls.h>
 
-#include "iree/base/alignment.h"
-
 // TODO: This should be cluster local.
 static struct worker_metadata_t {
   atomic_uint workers_waiting;
@@ -67,10 +69,17 @@ int quidditch_dispatch_enter_worker_loop() {
   while (!worker_metadata.exit) {
     park_worker();
     if (worker_metadata.exit) break;
-
+    
+    read_csr(mcycle);
     if (configuredKernel(configuredEnvironment, configuredDispatchState,
                          &configuredWorkgroupState[snrt_cluster_core_idx()]))
       error = true;
+
+    // Required to make sure that we only read the mcycle once the FPU has
+    // actually finished. Otherwise, we are measuring cycles that are too short!
+    // This is only required for measurement, not in real programs.
+    snrt_fpu_fence();
+    read_csr(mcycle);
   }
 
   snrt_interrupt_disable(IRQ_M_CLUSTER);
diff --git a/runtime/samples/nsnet2/CMakeLists.txt b/runtime/samples/nsnet2/CMakeLists.txt
index 5c2904c..fe36ded 100644
--- a/runtime/samples/nsnet2/CMakeLists.txt
+++ b/runtime/samples/nsnet2/CMakeLists.txt
@@ -9,4 +9,5 @@ target_link_libraries(
     samples_util
     nsnet2_module
     snRuntime
+    Quidditch::dispatch::dispatch
 )
diff --git a/runtime/samples/nsnet2/main.c b/runtime/samples/nsnet2/main.c
index b77da64..7173a05 100644
--- a/runtime/samples/nsnet2/main.c
+++ b/runtime/samples/nsnet2/main.c
@@ -1,9 +1,13 @@
+#include <Quidditch/dispatch/dispatch.h>
+
 #include <nsnet2.h>
 #include <nsnet2_module.h>
 #include <team_decls.h>
 #include <util/run_model.h>
 
 int main() {
+  if (!snrt_is_dm_core()) return quidditch_dispatch_enter_worker_loop();
+
   float data[161];
 
   for (int i = 0; i < IREE_ARRAYSIZE(data); i++) {
diff --git a/runtime/samples/util/run_model.c b/runtime/samples/util/run_model.c
index ea6bb4e..7a5aca6 100644
--- a/runtime/samples/util/run_model.c
+++ b/runtime/samples/util/run_model.c
@@ -71,12 +71,6 @@ static iree_status_t setup_instance_and_device(
 }
 
 iree_status_t run_model(const model_config_t* config) {
-  if (!snrt_is_dm_core()) {
-    int ret = quidditch_dispatch_enter_worker_loop();
-    if (!ret) return iree_ok_status();
-    return iree_make_status(IREE_STATUS_UNKNOWN);
-  }
-
   iree_allocator_t host_allocator = iree_allocator_system();
 
   iree_vm_instance_t* vmInstance;
diff --git a/runtime/samples/vec_multiply/CMakeLists.txt b/runtime/samples/vec_multiply/CMakeLists.txt
index e081c5a..0f8bd22 100644
--- a/runtime/samples/vec_multiply/CMakeLists.txt
+++ b/runtime/samples/vec_multiply/CMakeLists.txt
@@ -7,5 +7,6 @@ target_link_libraries(
     samples_util
     simple_add_module
     snRuntime
+    Quidditch::dispatch::dispatch
 )
 
diff --git a/runtime/samples/vec_multiply/main.c b/runtime/samples/vec_multiply/main.c
index 3116429..4e20342 100644
--- a/runtime/samples/vec_multiply/main.c
+++ b/runtime/samples/vec_multiply/main.c
@@ -1,3 +1,5 @@
+#include <Quidditch/dispatch/dispatch.h>
+
 #include <simple_add.h>
 #include <simple_add_module.h>
 #include <team_decls.h>
@@ -5,6 +7,7 @@
 
 int main() {
   double data[4];
+  if (!snrt_is_dm_core()) return quidditch_dispatch_enter_worker_loop();
 
   for (int i = 0; i < IREE_ARRAYSIZE(data); i++) {
     data[i] = (i + 1);
diff --git a/runtime/snitch_cluster/api/ssr_decls.h b/runtime/snitch_cluster/api/ssr_decls.h
new file mode 100644
index 0000000..55b9003
--- /dev/null
+++ b/runtime/snitch_cluster/api/ssr_decls.h
@@ -0,0 +1,4 @@
+
+#pragma once
+
+void snrt_fpu_fence();
diff --git a/runtime/snitch_cluster/rtl/src/snrt.c b/runtime/snitch_cluster/rtl/src/snrt.c
index c25de2a..dd63f50 100644
--- a/runtime/snitch_cluster/rtl/src/snrt.c
+++ b/runtime/snitch_cluster/rtl/src/snrt.c
@@ -10,8 +10,11 @@
 #include "riscv.c"
 #include "snitch_cluster_memory.c"
 #include "snitch_cluster_start.c"
+#include "ssr.h"
 #include "stack_decls.h"
 #include "sync.c"
 #include "team.c"
 
+extern void snrt_fpu_fence();
+
 uint32_t snrt_get_stack_size_per_core() { return 1 << SNRT_LOG2_STACK_SIZE; }