diff --git a/tensorflow/core/lib/jpeg/jpeg_mem.cc b/tensorflow/core/lib/jpeg/jpeg_mem.cc
index 56c247560b0..426120b2100 100644
--- a/tensorflow/core/lib/jpeg/jpeg_mem.cc
+++ b/tensorflow/core/lib/jpeg/jpeg_mem.cc
@@ -139,11 +139,8 @@ uint8* UncompressLow(const void* srcdata, FewerArgsForCompiler* argball) {
   cinfo.do_fancy_upsampling = boolean(flags.fancy_upscaling);
   cinfo.scale_num = 1;
   cinfo.scale_denom = ratio;
-  // Activating this has a quality/speed trade-off implication.
-  // However, most JPEGs are already compressed, and so the faster IDCT
-  // should have no effect on training. The fast setting speeds up training on the
-  // GPU, e.g. by about 30% for smaller networks such as AlexNet.
-  cinfo.dct_method = JDCT_IFAST;
+  // Activating this has a quality/speed trade-off implication:
+  // cinfo.dct_method = JDCT_IFAST;
 
   jpeg_start_decompress(&cinfo);
 
diff --git a/tensorflow/core/lib/jpeg/jpeg_mem_unittest.cc b/tensorflow/core/lib/jpeg/jpeg_mem_unittest.cc
index fc3961ad6ae..d0b4bef5d47 100644
--- a/tensorflow/core/lib/jpeg/jpeg_mem_unittest.cc
+++ b/tensorflow/core/lib/jpeg/jpeg_mem_unittest.cc
@@ -165,7 +165,7 @@ TEST(JpegMemTest, Jpeg2) {
     // Compare the two images
     const int totalerr = ComputeSumAbsoluteDifference(
         imgdata1.get(), refdata1.get(), in_w, in_h, stride1, stride1);
-    CHECK_LE(totalerr, 120000);
+    CHECK_LE(totalerr, 85000);
   }
 
   // check the second image too. Should be bitwise identical to the first.
diff --git a/tensorflow/python/ops/image_ops_test.py b/tensorflow/python/ops/image_ops_test.py
index ef9d4706d18..b47ecd2f8fb 100644
--- a/tensorflow/python/ops/image_ops_test.py
+++ b/tensorflow/python/ops/image_ops_test.py
@@ -1803,10 +1803,10 @@ class JpegTest(test_util.TensorFlowTestCase):
       jpeg0, image0, image1, image2 = sess.run([jpeg0, image0, image1, image2])
 
       # The decoded-encoded image should be similar to the input
-      self.assertLess(self.averageError(image0, image1), 0.7)
+      self.assertLess(self.averageError(image0, image1), 0.6)
 
       # We should be very close to a fixpoint
-      self.assertLess(self.averageError(image1, image2), 0.6)
+      self.assertLess(self.averageError(image1, image2), 0.02)
 
       # Smooth ramps compress well (input size is 153600)
       self.assertGreaterEqual(len(jpeg0), 5000)
diff --git a/tensorflow/tensorflow.bzl b/tensorflow/tensorflow.bzl
index 388dba33b53..31b4461dad6 100644
--- a/tensorflow/tensorflow.bzl
+++ b/tensorflow/tensorflow.bzl
@@ -416,7 +416,7 @@ def _cuda_copts():
         "@local_config_cuda//cuda:using_nvcc": (
             common_cuda_opts +
             [
-                "-nvcc_options=expt-relaxed-constexpr",
+                "-nvcc_options=relaxed-constexpr",
                 "-nvcc_options=ftz=true",
             ]
         ),