diff --git a/src/dev/amdgpu/sdma_engine.cc b/src/dev/amdgpu/sdma_engine.cc
index 582f128f82..6955837a09 100644
--- a/src/dev/amdgpu/sdma_engine.cc
+++ b/src/dev/amdgpu/sdma_engine.cc
@@ -38,6 +38,7 @@
 #include "dev/amdgpu/interrupt_handler.hh"
 #include "dev/amdgpu/sdma_commands.hh"
 #include "dev/amdgpu/sdma_mmio.hh"
+#include "gpu-compute/gpu_command_processor.hh"
 #include "mem/packet.hh"
 #include "mem/packet_access.hh"
 #include "params/SDMAEngine.hh"
@@ -653,11 +654,29 @@ SDMAEngine::writeDone(SDMAQueue *q, sdmaWrite *pkt, uint32_t *dmaBuffer)
 {
     DPRINTF(SDMAEngine, "Write packet completed to %p, %d dwords\n",
             pkt->dest, pkt->count);
-    delete [] dmaBuffer;
+
+    auto cleanup_cb = new EventFunctionWrapper(
+        [ = ]{ writeCleanup(dmaBuffer); }, name());
+
+    auto system_ptr = gpuDevice->CP()->system();
+    if (!system_ptr->isAtomicMode()) {
+        warn_once("SDMA cleanup assumes 2000 tick timing for completion."
+                " This has not been tested in timing mode\n");
+    }
+
+    // Only 2000 ticks should be necessary, but add additional padding.
+    schedule(cleanup_cb, curTick() + 10000);
+
     delete pkt;
     decodeNext(q);
 }
 
+void
+SDMAEngine::writeCleanup(uint32_t *dmaBuffer)
+{
+    delete [] dmaBuffer;
+}
+
 /* Implements a copy packet. */
 void
 SDMAEngine::copy(SDMAQueue *q, sdmaCopy *pkt)
@@ -747,6 +766,7 @@ SDMAEngine::copyReadData(SDMAQueue *q, sdmaCopy *pkt, uint8_t *dmaBuffer)
             buffer_ptr += gen.size();
         }
     } else {
+        DPRINTF(SDMAEngine, "Copying to host address %#lx\n", pkt->dest);
         auto cb = new DmaVirtCallback<uint64_t>(
             [ = ] (const uint64_t &) { copyDone(q, pkt, dmaBuffer); });
         dmaWriteVirt(pkt->dest, pkt->count, cb, (void *)dmaBuffer);
@@ -772,11 +792,29 @@ SDMAEngine::copyDone(SDMAQueue *q, sdmaCopy *pkt, uint8_t *dmaBuffer)
 {
     DPRINTF(SDMAEngine, "Copy completed to %p, %d dwords\n",
             pkt->dest, pkt->count);
-    delete [] dmaBuffer;
+
+    auto cleanup_cb = new EventFunctionWrapper(
+        [ = ]{ copyCleanup(dmaBuffer); }, name());
+
+    auto system_ptr = gpuDevice->CP()->system();
+    if (!system_ptr->isAtomicMode()) {
+        warn_once("SDMA cleanup assumes 2000 tick timing for completion."
+                " This has not been tested in timing mode\n");
+    }
+
+    // Only 2000 ticks should be necessary, but add additional padding.
+    schedule(cleanup_cb, curTick() + 10000);
+
     delete pkt;
     decodeNext(q);
 }
 
+void
+SDMAEngine::copyCleanup(uint8_t *dmaBuffer)
+{
+    delete [] dmaBuffer;
+}
+
 /* Implements an indirect buffer packet. */
 void
 SDMAEngine::indirectBuffer(SDMAQueue *q, sdmaIndirectBuffer *pkt)
@@ -1020,11 +1058,28 @@ SDMAEngine::ptePdeDone(SDMAQueue *q, sdmaPtePde *pkt, uint64_t *dmaBuffer)
     DPRINTF(SDMAEngine, "PtePde packet completed to %p, %d 2dwords\n",
             pkt->dest, pkt->count);
 
-    delete [] dmaBuffer;
+    auto cleanup_cb = new EventFunctionWrapper(
+        [ = ]{ ptePdeCleanup(dmaBuffer); }, name());
+
+    auto system_ptr = gpuDevice->CP()->system();
+    if (!system_ptr->isAtomicMode()) {
+        warn_once("SDMA cleanup assumes 2000 tick timing for completion."
+                " This has not been tested in timing mode\n");
+    }
+
+    // Only 2000 ticks should be necessary, but add additional padding.
+    schedule(cleanup_cb, curTick() + 10000);
+
     delete pkt;
     decodeNext(q);
 }
 
+void
+SDMAEngine::ptePdeCleanup(uint64_t *dmaBuffer)
+{
+    delete [] dmaBuffer;
+}
+
 void
 SDMAEngine::atomic(SDMAQueue *q, sdmaAtomicHeader *header, sdmaAtomic *pkt)
 {
diff --git a/src/dev/amdgpu/sdma_engine.hh b/src/dev/amdgpu/sdma_engine.hh
index 9407b97d73..d5fe646fac 100644
--- a/src/dev/amdgpu/sdma_engine.hh
+++ b/src/dev/amdgpu/sdma_engine.hh
@@ -227,9 +227,11 @@ class SDMAEngine : public DmaVirtDevice
     void write(SDMAQueue *q, sdmaWrite *pkt);
     void writeReadData(SDMAQueue *q, sdmaWrite *pkt, uint32_t *dmaBuffer);
     void writeDone(SDMAQueue *q, sdmaWrite *pkt, uint32_t *dmaBuffer);
+    void writeCleanup(uint32_t *dmaBuffer);
     void copy(SDMAQueue *q, sdmaCopy *pkt);
     void copyReadData(SDMAQueue *q, sdmaCopy *pkt, uint8_t *dmaBuffer);
     void copyDone(SDMAQueue *q, sdmaCopy *pkt, uint8_t *dmaBuffer);
+    void copyCleanup(uint8_t *dmaBuffer);
     void indirectBuffer(SDMAQueue *q, sdmaIndirectBuffer *pkt);
     void fence(SDMAQueue *q, sdmaFence *pkt);
     void fenceDone(SDMAQueue *q, sdmaFence *pkt);
@@ -243,6 +245,7 @@ class SDMAEngine : public DmaVirtDevice
     bool pollRegMemFunc(uint32_t value, uint32_t reference, uint32_t func);
     void ptePde(SDMAQueue *q, sdmaPtePde *pkt);
     void ptePdeDone(SDMAQueue *q, sdmaPtePde *pkt, uint64_t *dmaBuffer);
+    void ptePdeCleanup(uint64_t *dmaBuffer);
     void atomic(SDMAQueue *q, sdmaAtomicHeader *header, sdmaAtomic *pkt);
     void atomicData(SDMAQueue *q, sdmaAtomicHeader *header, sdmaAtomic *pkt,
                     uint64_t *dmaBuffer);