diff --git a/tensorflow/compiler/xla/service/hlo_rematerialization.cc b/tensorflow/compiler/xla/service/hlo_rematerialization.cc
index 21be4216469..bfc6769660a 100644
--- a/tensorflow/compiler/xla/service/hlo_rematerialization.cc
+++ b/tensorflow/compiler/xla/service/hlo_rematerialization.cc
@@ -1648,6 +1648,8 @@ StatusOr<bool> HloRematerialization::RematerializeComputation(
       } else {
         // Found a valid block. Reset to start looking for single instructions
         // again.
+        max_rematerialized_block_size_ =
+            std::max(max_rematerialized_block_size_, max_block_size);
         changed = true;
         min_block_size = 1;
         max_block_size = 1;
diff --git a/tensorflow/compiler/xla/service/hlo_rematerialization.h b/tensorflow/compiler/xla/service/hlo_rematerialization.h
index d1c4b8b5e7b..72221fa8a32 100644
--- a/tensorflow/compiler/xla/service/hlo_rematerialization.h
+++ b/tensorflow/compiler/xla/service/hlo_rematerialization.h
@@ -180,6 +180,10 @@ class HloRematerialization : public HloModulePass {
   // dead. Hence, no net instructions were added.
   int64 net_instructions_added_ = 0;
 
+  // Size of the largest block that has been rematerialized. This is actually an
+  // upper bound (within a factor of 2) on the block size.
+  int max_rematerialized_block_size_ = 0;
+
   RematerializationMode mode_;
 };