diff --git a/projects/rocr-runtime/include/hsakmttypes.h b/projects/rocr-runtime/include/hsakmttypes.h
index 7a7dc2ad10..b821268c50 100644
--- a/projects/rocr-runtime/include/hsakmttypes.h
+++ b/projects/rocr-runtime/include/hsakmttypes.h
@@ -541,7 +541,8 @@ typedef struct _HsaMemFlags
 					    // and optimal alignment requirements
             unsigned int FixedAddress : 1; // Allocate memory at specified virtual address. Fail if address is not free.
             unsigned int NoNUMABind:    1; // Don't bind system memory to a specific NUMA node
-            unsigned int Reserved    : 15;
+            unsigned int Uncached:      1; // Caching flag for fine-grained memory on A+A HW platform
+            unsigned int Reserved    : 14;
 
         } ui32;
         HSAuint32 Value;
diff --git a/projects/rocr-runtime/include/linux/kfd_ioctl.h b/projects/rocr-runtime/include/linux/kfd_ioctl.h
index 6a7dd1f44a..aa00e919dd 100644
--- a/projects/rocr-runtime/include/linux/kfd_ioctl.h
+++ b/projects/rocr-runtime/include/linux/kfd_ioctl.h
@@ -479,6 +479,7 @@ struct kfd_ioctl_acquire_vm_args {
 #define KFD_IOC_ALLOC_MEM_FLAGS_NO_SUBSTITUTE	(1 << 28)
 #define KFD_IOC_ALLOC_MEM_FLAGS_AQL_QUEUE_MEM	(1 << 27)
 #define KFD_IOC_ALLOC_MEM_FLAGS_COHERENT	(1 << 26)
+#define KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED	(1 << 25)
 
 /* Allocate memory for later SVM (shared virtual memory) mapping.
  *
diff --git a/projects/rocr-runtime/src/fmm.c b/projects/rocr-runtime/src/fmm.c
index 6867c69599..5fe6518ea6 100644
--- a/projects/rocr-runtime/src/fmm.c
+++ b/projects/rocr-runtime/src/fmm.c
@@ -1334,8 +1334,11 @@ void *fmm_allocate_device(uint32_t gpu_id, void *address, uint64_t MemorySizeInB
 		aperture = &gpu_mem[gpu_mem_id].gpuvm_aperture;
 	}
 
-	if (!flags.ui32.CoarseGrain || svm.disable_cache)
+	if (!flags.ui32.CoarseGrain || svm.disable_cache) {
 		ioc_flags |= KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
+		if (flags.ui32.Uncached)
+			ioc_flags |= KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED;
+	}
 
 	mem = __fmm_allocate_device(gpu_id, address, size, aperture, &mmap_offset,
 				    ioc_flags, &vm_obj);
@@ -1540,8 +1543,11 @@ static void *fmm_allocate_host_gpu(uint32_t node_id, void *address,
 	else
 		aperture = svm.dgpu_alt_aperture; /* always coherent */
 
-	if (!flags.ui32.CoarseGrain || svm.disable_cache)
+	if (!flags.ui32.CoarseGrain || svm.disable_cache) {
 		ioc_flags |= KFD_IOC_ALLOC_MEM_FLAGS_COHERENT;
+		if (flags.ui32.Uncached)
+			ioc_flags |= KFD_IOC_ALLOC_MEM_FLAGS_UNCACHED;
+	}
 	ioc_flags |= fmm_translate_hsa_to_ioc_flags(flags);
 
 	if (flags.ui32.AQLQueueMemory)