diff --git a/csrc/custom_all_reduce.cuh b/csrc/custom_all_reduce.cuh
index 632b579c55afa..7b997f2a58ae1 100644
--- a/csrc/custom_all_reduce.cuh
+++ b/csrc/custom_all_reduce.cuh
@@ -175,7 +175,7 @@ DINLINE void multi_gpu_barrier(const RankSignals& sg, Signal* self_sg,
         &sg.signals[threadIdx.x]->peer_counter[val % 2][blockIdx.x][rank];
     auto self_counter_ptr =
         &self_sg->peer_counter[val % 2][blockIdx.x][threadIdx.x];
-    if constexpr (need_fence) {
+    if constexpr (false) {
       st_flag_release(peer_counter_ptr, val);
       while (ld_flag_acquire(self_counter_ptr) != val);
     } else {
@@ -240,6 +240,17 @@ __global__ void __launch_bounds__(512, 1)
     tmps[i] = get_tmp_buf<P>(sg.signals[target]);
   }
   auto tmp_out = tmps[0];
+// Uncomment to populate cache of peer values, which are currently in a indeterminate state
+//   for (int idx = tid; idx < largest_part; idx += stride) {
+// #pragma unroll
+//     for (int i = 0; i < ngpus; i++) {
+//       int gather_from_rank = ((rank + i) % ngpus);
+//       if (gather_from_rank == ngpus - 1 || idx < part) {
+//         int dst_idx = gather_from_rank * part + idx;
+//         ((P*)result)[dst_idx] = tmps[i][idx];
+//       }
+//     }
+//   }
   multi_gpu_barrier<ngpus, true>(sg, self_sg, rank);
   // stage 1: reduce scatter
   for (int idx = start + tid; idx < end; idx += stride) {
@@ -258,7 +269,18 @@ __global__ void __launch_bounds__(512, 1)
       int gather_from_rank = ((rank + i) % ngpus);
       if (gather_from_rank == ngpus - 1 || idx < part) {
         int dst_idx = gather_from_rank * part + idx;
-        ((P*)result)[dst_idx] = tmps[i][idx];
+        uint4 tmp;
+        // note: relaxed.sys also works
+        asm volatile (
+            "ld.volatile.global.v4.u32 {%0, %1, %2, %3}, [%4];\n"
+            : "=r"(tmp.x), "=r"(tmp.y), "=r"(tmp.z), "=r"(tmp.w)
+            : "l"(&tmps[i][idx])
+        );
+        // asm volatile (
+        //     "st.global.v4.u32 [%4], {%0, %1, %2, %3};\n"
+        //     :: "r"(tmp.x), "r"(tmp.y), "r"(tmp.z), "r"(tmp.w), "l"(&((P*)result)[dst_idx])
+        // );
+        ((P*)result)[dst_idx] = *reinterpret_cast<P*>(&tmp);
       }
     }
   }