diff --git a/icd/intel/cmd.c b/icd/intel/cmd.c
index 5a888ae..c2237b9 100644
--- a/icd/intel/cmd.c
+++ b/icd/intel/cmd.c
@@ -147,7 +147,7 @@
     if (cmd->flags != flags || !bo_size) {
         cmd->flags = flags;
 
-        bo_size = cmd->dev->gpu->batch_buffer_size;
+        bo_size = cmd->dev->gpu->max_batch_buffer_size;
         if (flags & XGL_CMD_BUFFER_OPTIMIZE_GPU_SMALL_BATCH_BIT)
             bo_size /= 2;
 
diff --git a/icd/intel/gpu.c b/icd/intel/gpu.c
index 4372929..e2aa305 100644
--- a/icd/intel/gpu.c
+++ b/icd/intel/gpu.c
@@ -115,7 +115,11 @@
     gpu->gen_opaque = gen;
 
     /* 8192 dwords */
-    gpu->batch_buffer_size = sizeof(uint32_t) * 8192;
+    gpu->max_batch_buffer_size = sizeof(uint32_t) * 8192;
+
+    /* the winsys is prepared for one reloc every two dwords, then minus 2 */
+    gpu->batch_buffer_reloc_count =
+        gpu->max_batch_buffer_size / sizeof(uint32_t) / 2 - 2;
 
     gpu->fd_internal = -1;
     gpu->fd = -1;
@@ -231,9 +235,7 @@
     memcpy(props->gpuName, name, name_len);
     props->gpuName[name_len] = '\0';
 
-    /* the winsys is prepared for one reloc every two dwords, then minus 2 */
-    props->maxMemRefsPerSubmission =
-        gpu->batch_buffer_size / sizeof(uint32_t) / 2 - 2;
+    props->maxMemRefsPerSubmission = gpu->batch_buffer_reloc_count;
 
     props->virtualMemPageSize = 4096;
 
diff --git a/icd/intel/gpu.h b/icd/intel/gpu.h
index fac16bc..e5d8db3 100644
--- a/icd/intel/gpu.h
+++ b/icd/intel/gpu.h
@@ -50,7 +50,8 @@
     char *path;         /* path to the render or legacy node, or NULL */
     int gen_opaque;     /* always read with intel_gpu_gen() */
 
-    int batch_buffer_size;
+    XGL_SIZE max_batch_buffer_size;
+    XGL_UINT batch_buffer_reloc_count;
 
     /*
      * The enabled hardware features could be limited by the kernel.  This