diff --git a/opal/mca/btl/ugni/btl_ugni_atomic.c b/opal/mca/btl/ugni/btl_ugni_atomic.c
index 981bc759ee..3c62670da8 100644
--- a/opal/mca/btl/ugni/btl_ugni_atomic.c
+++ b/opal/mca/btl/ugni/btl_ugni_atomic.c
@@ -1,6 +1,6 @@
 /* -*- Mode: C; c-basic-offset:4 ; indent-tabs-mode:nil -*- */
 /*
- * Copyright (c) 2014      Los Alamos National Security, LLC. All rights
+ * Copyright (c) 2014-2016 Los Alamos National Security, LLC. All rights
  *                         reserved.
  * $COPYRIGHT$
  *
@@ -11,18 +11,66 @@
 
 #include "btl_ugni_rdma.h"
 
-static gni_fma_cmd_type_t famo_cmds[] = {
-    [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC_FADD,
-    [MCA_BTL_ATOMIC_AND] = GNI_FMA_ATOMIC_FAND,
-    [MCA_BTL_ATOMIC_OR] = GNI_FMA_ATOMIC_FOR,
-    [MCA_BTL_ATOMIC_XOR] = GNI_FMA_ATOMIC_FXOR,
+static gni_fma_cmd_type_t amo_cmds[][MCA_BTL_ATOMIC_LAST] = {
+    [OPAL_INT32] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC2_IADD_S,
+        [MCA_BTL_ATOMIC_LAND] = GNI_FMA_ATOMIC2_AND_S,
+        [MCA_BTL_ATOMIC_LOR] = GNI_FMA_ATOMIC2_OR_S,
+        [MCA_BTL_ATOMIC_LXOR] = GNI_FMA_ATOMIC2_XOR_S,
+        [MCA_BTL_ATOMIC_SWAP] = GNI_FMA_ATOMIC2_SWAP_S,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_IMIN_S,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_IMAX_S,
+    },
+    [OPAL_INT64] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC_ADD,
+        [MCA_BTL_ATOMIC_AND] = GNI_FMA_ATOMIC_AND,
+        [MCA_BTL_ATOMIC_OR] = GNI_FMA_ATOMIC_OR,
+        [MCA_BTL_ATOMIC_XOR] = GNI_FMA_ATOMIC_XOR,
+        [MCA_BTL_ATOMIC_SWAP] = GNI_FMA_ATOMIC2_SWAP,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_IMIN,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_IMAX,
+    },
+    [OPAL_FLOAT] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC2_FPADD_S,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_FPMIN_S,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_FPMAX_S,
+    },
+    [OPAL_DOUBLE] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC2_FPADD,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_FPMIN,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_FPMAX,
+    },
 };
 
-static gni_fma_cmd_type_t amo_cmds[] = {
-    [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC_ADD,
-    [MCA_BTL_ATOMIC_AND] = GNI_FMA_ATOMIC_AND,
-    [MCA_BTL_ATOMIC_OR] = GNI_FMA_ATOMIC_OR,
-    [MCA_BTL_ATOMIC_XOR] = GNI_FMA_ATOMIC_XOR,
+static gni_fma_cmd_type_t famo_cmds[][MCA_BTL_ATOMIC_LAST] = {
+    [OPAL_INT32] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC2_FIADD_S,
+        [MCA_BTL_ATOMIC_LAND] = GNI_FMA_ATOMIC2_FAND_S,
+        [MCA_BTL_ATOMIC_LOR] = GNI_FMA_ATOMIC2_FOR_S,
+        [MCA_BTL_ATOMIC_LXOR] = GNI_FMA_ATOMIC2_FXOR_S,
+        [MCA_BTL_ATOMIC_SWAP] = GNI_FMA_ATOMIC2_FSWAP_S,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_FIMIN_S,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_FIMAX_S,
+    },
+    [OPAL_INT64] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC_FADD,
+        [MCA_BTL_ATOMIC_AND] = GNI_FMA_ATOMIC_FAND,
+        [MCA_BTL_ATOMIC_OR] = GNI_FMA_ATOMIC_FOR,
+        [MCA_BTL_ATOMIC_XOR] = GNI_FMA_ATOMIC_FXOR,
+        [MCA_BTL_ATOMIC_SWAP] = GNI_FMA_ATOMIC2_FSWAP,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_FIMIN,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_FIMAX,
+    },
+    [OPAL_FLOAT] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC2_FFPADD_S,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_FFPMIN_S,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_FFPMAX_S,
+    },
+    [OPAL_DOUBLE] = {
+        [MCA_BTL_ATOMIC_ADD] = GNI_FMA_ATOMIC2_FFPADD,
+        [MCA_BTL_ATOMIC_MIN] = GNI_FMA_ATOMIC2_FFPMIN,
+        [MCA_BTL_ATOMIC_MAX] = GNI_FMA_ATOMIC2_FFPMAX,
+    },
 };
 
 int mca_btl_ugni_aop (struct mca_btl_base_module_t *btl, struct mca_btl_base_endpoint_t *endpoint,
@@ -32,7 +80,20 @@ int mca_btl_ugni_aop (struct mca_btl_base_module_t *btl, struct mca_btl_base_end
 {
     gni_mem_handle_t dummy = {0, 0};
     mca_btl_ugni_post_descriptor_t *post_desc;
-    int rc;
+    int gni_op, rc, type;
+    size_t size;
+
+    size = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? 4 : 8;
+    if (MCA_BTL_ATOMIC_FLAG_FLOAT & flags) {
+        type = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? OPAL_FLOAT : OPAL_DOUBLE;
+    } else {
+        type = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? OPAL_INT32 : OPAL_INT64;
+    }
+
+    gni_op = amo_cmds[type][op];
+    if (0 == gni_op) {
+        return OPAL_ERR_NOT_SUPPORTED;
+    }
 
     rc = mca_btl_ugni_check_endpoint_state_rdma (endpoint);
     if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
@@ -45,8 +106,8 @@ int mca_btl_ugni_aop (struct mca_btl_base_module_t *btl, struct mca_btl_base_end
     }
 
     init_gni_post_desc (&post_desc->desc, order, GNI_POST_AMO, 0, dummy, remote_address,
-                        remote_handle->gni_handle, 8, 0);
-    post_desc->desc.base.amo_cmd = amo_cmds[op];
+                        remote_handle->gni_handle, size, 0);
+    post_desc->desc.base.amo_cmd = gni_op;
 
     post_desc->desc.base.first_operand = operand;
 
@@ -54,6 +115,10 @@ int mca_btl_ugni_aop (struct mca_btl_base_module_t *btl, struct mca_btl_base_end
     rc = GNI_PostFma (endpoint->rdma_ep_handle, &post_desc->desc.base);
     OPAL_THREAD_UNLOCK(&endpoint->btl->device->dev_lock);
     if (GNI_RC_SUCCESS != rc) {
+        mca_btl_ugni_return_post_descriptor (endpoint->btl, post_desc);
+        if (GNI_RC_ILLEGAL_OP == rc) {
+            return OPAL_ERR_NOT_SUPPORTED;
+        }
         return OPAL_ERR_OUT_OF_RESOURCE;
     }
 
@@ -67,7 +132,20 @@ int mca_btl_ugni_afop (struct mca_btl_base_module_t *btl, struct mca_btl_base_en
                        void *cbcontext, void *cbdata)
 {
     mca_btl_ugni_post_descriptor_t *post_desc;
-    int rc;
+    int gni_op, rc, type;
+    size_t size;
+
+    size = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? 4 : 8;
+    if (MCA_BTL_ATOMIC_FLAG_FLOAT & flags) {
+        type = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? OPAL_FLOAT : OPAL_DOUBLE;
+    } else {
+        type = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? OPAL_INT32 : OPAL_INT64;
+    }
+
+    gni_op = famo_cmds[type][op];
+    if (0 == gni_op) {
+        return OPAL_ERR_NOT_SUPPORTED;
+    }
 
     rc = mca_btl_ugni_check_endpoint_state_rdma (endpoint);
     if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
@@ -81,8 +159,8 @@ int mca_btl_ugni_afop (struct mca_btl_base_module_t *btl, struct mca_btl_base_en
 
 
     init_gni_post_desc (&post_desc->desc, order, GNI_POST_AMO, (intptr_t) local_address, local_handle->gni_handle,
-                        remote_address, remote_handle->gni_handle, 8, 0);
-    post_desc->desc.base.amo_cmd = famo_cmds[op];
+                        remote_address, remote_handle->gni_handle, size, 0);
+    post_desc->desc.base.amo_cmd = gni_op;
 
     post_desc->desc.base.first_operand = operand;
 
@@ -91,6 +169,9 @@ int mca_btl_ugni_afop (struct mca_btl_base_module_t *btl, struct mca_btl_base_en
     OPAL_THREAD_UNLOCK(&endpoint->btl->device->dev_lock);
     if (GNI_RC_SUCCESS != rc) {
         mca_btl_ugni_return_post_descriptor (endpoint->btl, post_desc);
+        if (GNI_RC_ILLEGAL_OP == rc) {
+            return OPAL_ERR_NOT_SUPPORTED;
+        }
         return OPAL_ERR_OUT_OF_RESOURCE;
     }
 
@@ -103,7 +184,11 @@ int mca_btl_ugni_acswap (struct mca_btl_base_module_t *btl, struct mca_btl_base_
                          int order, mca_btl_base_rdma_completion_fn_t cbfunc, void *cbcontext, void *cbdata)
 {
     mca_btl_ugni_post_descriptor_t *post_desc;
-    int rc;
+    int gni_op, rc;
+    size_t size;
+
+    gni_op = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? GNI_FMA_ATOMIC2_CSWAP_S : GNI_FMA_ATOMIC_CSWAP;
+    size = (MCA_BTL_ATOMIC_FLAG_32BIT & flags) ? 4 : 8;
 
     rc = mca_btl_ugni_check_endpoint_state_rdma (endpoint);
     if (OPAL_UNLIKELY(OPAL_SUCCESS != rc)) {
@@ -117,8 +202,8 @@ int mca_btl_ugni_acswap (struct mca_btl_base_module_t *btl, struct mca_btl_base_
 
 
     init_gni_post_desc (&post_desc->desc, order, GNI_POST_AMO, (intptr_t) local_address, local_handle->gni_handle,
-                        remote_address, remote_handle->gni_handle, 8, 0);
-    post_desc->desc.base.amo_cmd = GNI_FMA_ATOMIC_CSWAP;
+                        remote_address, remote_handle->gni_handle, size, 0);
+    post_desc->desc.base.amo_cmd = gni_op;
 
     post_desc->desc.base.first_operand = compare;
     post_desc->desc.base.second_operand = value;
diff --git a/opal/mca/btl/ugni/btl_ugni_component.c b/opal/mca/btl/ugni/btl_ugni_component.c
index 04076a16d6..e1351f40d0 100644
--- a/opal/mca/btl/ugni/btl_ugni_component.c
+++ b/opal/mca/btl/ugni/btl_ugni_component.c
@@ -291,6 +291,13 @@ btl_ugni_component_register(void)
         MCA_BTL_ATOMIC_SUPPORTS_AND | MCA_BTL_ATOMIC_SUPPORTS_OR | MCA_BTL_ATOMIC_SUPPORTS_XOR |
         MCA_BTL_ATOMIC_SUPPORTS_CSWAP;
 
+    if (GNI_DEVICE_ARIES == device_type) {
+        /* aries supports additional atomic operations */
+        mca_btl_ugni_module.super.btl_atomic_flags |= MCA_BTL_ATOMIC_SUPPORTS_MIN | MCA_BTL_ATOMIC_SUPPORTS_MAX |
+            MCA_BTL_ATOMIC_SUPPORTS_LAND | MCA_BTL_ATOMIC_SUPPORTS_LOR | MCA_BTL_ATOMIC_SUPPORTS_LXOR |
+            MCA_BTL_ATOMIC_SUPPORTS_32BIT | MCA_BTL_ATOMIC_SUPPORTS_FLOAT;
+    }
+
     mca_btl_ugni_module.super.btl_registration_handle_size = sizeof (mca_btl_base_registration_handle_t);
 
     mca_btl_ugni_module.super.btl_bandwidth = 40000; /* Mbs */