diff --git a/target/sim/sw/device/runtime/src/occamy_device.h b/target/sim/sw/device/runtime/src/occamy_device.h
index dcc1fbde..b769e8f4 100644
--- a/target/sim/sw/device/runtime/src/occamy_device.h
+++ b/target/sim/sw/device/runtime/src/occamy_device.h
@@ -55,14 +55,17 @@ inline void return_to_cva6(sync_t sync) {
             if (cnt == snrt_cluster_num()) {
 #endif
                 *((volatile uint32_t*)barrier_ptr) = 0;
-                // Interrupt the local host to signal the exit code (snitch by default only has the access to local domain)
-                set_host_sw_interrupt(0);
+                // Interrupt the local host to signal the exit code (snitch by
+                // default only has the access to local domain)
+                comm_buffer_t* comm_buffer = get_communication_buffer();
+                set_host_sw_interrupt(comm_buffer->chip_id);
             }
         }
     }
     // Otherwise assume cores are already synchronized and only
     // one core calls this function
     else {
-        set_host_sw_interrupt(0);
+        comm_buffer_t* comm_buffer = get_communication_buffer();
+        set_host_sw_interrupt(comm_buffer->chip_id);
     }
 }
diff --git a/target/sim/sw/device/runtime/src/occamy_start.c b/target/sim/sw/device/runtime/src/occamy_start.c
index 00fca72b..e8c33381 100644
--- a/target/sim/sw/device/runtime/src/occamy_start.c
+++ b/target/sim/sw/device/runtime/src/occamy_start.c
@@ -26,7 +26,10 @@ static inline void snrt_exit_default(int exit_code);
 static inline void snrt_exit(int exit_code) {
     snrt_exit_default(exit_code);
     // Interrupt the local host to signal the exit code (snitch by default only has the access to local domain)
-    if (snrt_global_core_idx() == 0) set_host_sw_interrupt(0);
+    if (snrt_global_core_idx() == 0) {
+        comm_buffer_t* comm_buffer = get_communication_buffer();
+        set_host_sw_interrupt(comm_buffer->chip_id);
+    }
 }
 
 #include "start.c"
diff --git a/target/sim/sw/host/apps/offload/src/offload.c b/target/sim/sw/host/apps/offload/src/offload.c
index 5dd4b24d..e2f7af1b 100644
--- a/target/sim/sw/host/apps/offload/src/offload.c
+++ b/target/sim/sw/host/apps/offload/src/offload.c
@@ -55,19 +55,19 @@ int main() {
     reset_and_ungate_quadrants_all(target_chip_id);
     // print_str(current_chip_address_prefix, "[Occamy] Snitch ungated. \r\n");
     deisolate_all(target_chip_id);
-    // print_str(current_chip_address_prefix, "[Occamy] Snitch deisolated. \r\n");
-    // Enable interrupts to receive notice of job termination
+    // print_str(current_chip_address_prefix, "[Occamy] Snitch deisolated.
+    // \r\n"); Enable interrupts to receive notice of job termination
     enable_sw_interrupts();
     // Program Snitch entry point and communication buffer
-    (*comm_buffer_ptr).lock = 0;
-    (*comm_buffer_ptr).chip_id = current_chip_id;
+    comm_buffer_ptr->lock = 0;
+    comm_buffer_ptr->chip_id = current_chip_id;
     program_snitches(target_chip_id, comm_buffer_ptr);
     // print_str(current_chip_address_prefix,
     //           "[Occamy] Snitch Jump Address Programmed. \r\n");
 
     // Compiler fence to ensure Snitch entry point is
     // programmed before Snitches are woken up
-    asm volatile("" ::: "memory");
+    asm volatile("fence.i" ::: "memory");
 
     print_str(current_chip_address_prefix,
               "[Occamy] Calling snitch cluster to execute the task \r\n");
diff --git a/target/sim/sw/shared/runtime/chip_id.h b/target/sim/sw/shared/runtime/chip_id.h
index 04928e84..52d6ab67 100644
--- a/target/sim/sw/shared/runtime/chip_id.h
+++ b/target/sim/sw/shared/runtime/chip_id.h
@@ -37,6 +37,10 @@ inline uint8_t *get_chip_baseaddress(uint8_t chip_id) {
 #endif
 }
 
+inline uint32_t get_chip_baseaddress_h(uint8_t chip_id) {
+    return (uint32_t)(chip_id << 8);
+}
+
 inline uint32_t get_current_chip_baseaddress_h() {
     uint32_t chip_id = get_current_chip_id();
     return (uint32_t)(chip_id << 8);
diff --git a/target/sim/sw/shared/runtime/heterogeneous_runtime.h b/target/sim/sw/shared/runtime/heterogeneous_runtime.h
index a33ab31a..96b3397d 100644
--- a/target/sim/sw/shared/runtime/heterogeneous_runtime.h
+++ b/target/sim/sw/shared/runtime/heterogeneous_runtime.h
@@ -22,26 +22,85 @@ typedef struct {
 /**************/
 
 inline static void set_host_sw_interrupt(uint8_t chip_id) {
+#if __riscv_xlen == 64
     uint32_t* msip_ptr =
         (uint32_t*)(((uintptr_t)clint_msip_ptr(0)) |
                     ((uintptr_t)get_chip_baseaddress(chip_id)));
     *msip_ptr = 1;
+#elif __riscv_xlen == 32
+    uint32_t* msip_ptr = clint_msip_ptr(0);
+    uint32_t target_addrh = get_chip_baseaddress_h(chip_id);
+    uint32_t current_addrh = get_current_chip_baseaddress_h();
+
+    register uint32_t reg_target_addrh asm("t0") = target_addrh;
+    register uint32_t reg_return_value asm("t1") = 1;
+    register uint32_t reg_msip_ptr asm("t2") = (uint32_t)msip_ptr;
+    register uint32_t reg_current_addrh asm("t3") = current_addrh;
+
+    asm volatile(
+        "csrw 0xbc0, t0;"
+        "sw t1, 0(t2);"
+        "csrw 0xbc0, t3;"
+        :
+        : "r"(reg_target_addrh), "r"(reg_return_value), "r"(reg_msip_ptr),
+          "r"(reg_current_addrh)
+        : "memory");
+#endif
 }
 
 inline void clear_host_sw_interrupt_unsafe(uint8_t chip_id) {
+#if __riscv_xlen == 64
     uint32_t* msip_ptr =
         (uint32_t*)(((uintptr_t)clint_msip_ptr(0)) |
                     ((uintptr_t)get_chip_baseaddress(chip_id)));
-
     *msip_ptr = 0;
+#elif __riscv_xlen == 32
+    uint32_t* msip_ptr = clint_msip_ptr(0);
+    uint32_t target_addrh = get_chip_baseaddress_h(chip_id);
+    uint32_t current_addrh = get_current_chip_baseaddress_h();
+
+    register uint32_t reg_target_addrh asm("t0") = target_addrh;
+    register uint32_t reg_return_value asm("t1") = 0;
+    register uint32_t reg_msip_ptr asm("t2") = (uint32_t)msip_ptr;
+    register uint32_t reg_current_addrh asm("t3") = current_addrh;
+
+    asm volatile(
+        "csrw 0xbc0, t0;"
+        "sw t1, 0(t2);"
+        "csrw 0xbc0, t3;"
+        :
+        : "r"(reg_target_addrh), "r"(reg_return_value), "r"(reg_msip_ptr),
+          "r"(reg_current_addrh)
+        : "memory");
+#endif
 }
 
 inline void wait_host_sw_interrupt_clear(uint8_t chip_id) {
+#if __riscv_xlen == 64
     uint32_t* msip_ptr =
         (uint32_t*)(((uintptr_t)clint_msip_ptr(0)) |
                     ((uintptr_t)get_chip_baseaddress(chip_id)));
-
     while (*msip_ptr);
+#elif __riscv_xlen == 32
+    uint32_t* msip_ptr = clint_msip_ptr(0);
+    uint32_t target_addrh = get_chip_baseaddress_h(chip_id);
+    uint32_t current_addrh = get_current_chip_baseaddress_h();
+
+    register uint32_t reg_target_addrh asm("t0") = target_addrh;
+    register uint32_t reg_value asm("t1");
+    register uint32_t reg_msip_ptr asm("t2") = (uint32_t)msip_ptr;
+    register uint32_t reg_current_addrh asm("t3") = current_addrh;
+
+    do {
+        asm volatile(
+            "csrw 0xbc0, t0;"
+            "lw t1, 0(t2);"
+            "csrw 0xbc0, t3;"
+            : "=r"(reg_value)
+            : "r"(reg_target_addrh), "r"(reg_msip_ptr), "r"(reg_current_addrh)
+            : "memory");
+    } while (reg_value);
+#endif
 }
 
 static inline void clear_host_sw_interrupt(uint8_t chip_id) {
diff --git a/target/sim_chip/apps/Makefile b/target/sim_chip/apps/Makefile
index 5a7bb53d..72a2846a 100644
--- a/target/sim_chip/apps/Makefile
+++ b/target/sim_chip/apps/Makefile
@@ -13,11 +13,11 @@ RISCV_READELF = $(CVA6_GCC_ROOT)/riscv64-unknown-elf-readelf
 ELFS = $(shell find $(TARGET)/sim/sw/host/apps -type f -name "*.elf")
 # Generate BIN filenames from the ELF filenames
 BINS = $(ELFS:.elf=.bin)
-$(info BINS: $(BINS))
 
-.PHONY: apps clean
+.PHONY: apps hex clean
 
 apps: $(BINS)
+	python3 bin2hex.py
 
 # Use full paths for the dependencies
 %.bin: %.elf
diff --git a/util/occamygen/occamy.py b/util/occamygen/occamy.py
index 8dc50948..5be51ffa 100644
--- a/util/occamygen/occamy.py
+++ b/util/occamygen/occamy.py
@@ -315,10 +315,10 @@ def am_connect_soc_wide_xbar_quad(am, am_soc_narrow_xbar, am_wide_xbar_quadrant_
                     "quadrant_{}_cluster_{}_tcdm".format(i, j),
                     clusters_tcdm_size[j+1],
                     *bases_cluster
-                ).attach_to(
-                    am_wide_xbar_quadrant_s1[i]
                 ).attach_to(
                     am_narrow_xbar_quadrant_s1[i]
+                ).attach_to(
+                    am_wide_xbar_quadrant_s1[i]
                 )
             )
 
@@ -330,11 +330,11 @@ def am_connect_soc_wide_xbar_quad(am, am_soc_narrow_xbar, am_wide_xbar_quadrant_
                     "quadrant_{}_cluster_{}_periph".format(i, j),
                     clusters_periph_size[j+1],
                     *bases_cluster
-                ).attach_to(
-                    am_wide_xbar_quadrant_s1[i]
                 ).attach_to(
                     am_narrow_xbar_quadrant_s1[i]
-                )
+                ) # .attach_to(
+                #     am_wide_xbar_quadrant_s1[i]
+                # )
             )
 
             bases_cluster = list()