diff --git a/sw/apps/transformer/src/transformer.h b/sw/apps/transformer/src/transformer.h
index 8b5bb7bf1..a19a56a96 100644
--- a/sw/apps/transformer/src/transformer.h
+++ b/sw/apps/transformer/src/transformer.h
@@ -942,17 +942,23 @@ static inline void transformer_layer_fp64(transformer_layer_fp64_t *const l) {
 
         // now we will add the partial results together
         // in a logarithmic reduction fashion
-        uint32_t cluster_offset = 0x40000;
+        uint32_t cl_offset = 0x40000;
         // num_levels: number of levels in the reduction tree
         int num_levels = (int)log2(num_cores);
         for (int level = 0; level < num_levels; level++) {
+            // determine whether the current core is an active core
+            uint32_t is_active = (compute_id % (1 << level)) == 0;
+            dump_id(level); // CSR 5
+            dump_idx(is_active); // CSR 6
+
             // every core adds its partial result to the result of the core
             // that is 2^level away from it, except for the last core after
             // the first iteration
             uint32_t src_core_id = compute_id;
             uint32_t dst_core_id = compute_id + (1 << level);
-            dump_idx(src_core_id);
-            dump_idx(dst_core_id);
+            // dump_idx(src_core_id);
+            // dump_idx(dst_core_id);
+        }
 
     }