diff --git a/onnxruntime/test/global_thread_pools/test_inference.cc b/onnxruntime/test/global_thread_pools/test_inference.cc
index 4772e7de2b..f553682975 100644
--- a/onnxruntime/test/global_thread_pools/test_inference.cc
+++ b/onnxruntime/test/global_thread_pools/test_inference.cc
@@ -55,9 +55,15 @@ static void RunSession(OrtAllocator& allocator, Ort::Session& session_object,
   // size_t total_len = type_info.GetElementCount();
   ASSERT_EQ(values_y.size(), static_cast<size_t>(5));
 
+// test inference is using onnxruntime_shared_lib_test_LIBS, so HasCudaEnvironment(800) isn't available
+#ifdef USE_CUDA
+  const float tolerance = 1e-5f;
+#else
+  const float tolerance = 1e-6f;
+#endif
   OutT* f = output_tensor->GetTensorMutableData<OutT>();
   for (size_t i = 0; i != static_cast<size_t>(5); ++i) {
-    ASSERT_NEAR(values_y[i], f[i], 1e-6f);
+    ASSERT_NEAR(values_y[i], f[i], tolerance);
   }
 }
 
diff --git a/onnxruntime/test/providers/cpu/model_tests.cc b/onnxruntime/test/providers/cpu/model_tests.cc
index 859e082716..8128c170c5 100644
--- a/onnxruntime/test/providers/cpu/model_tests.cc
+++ b/onnxruntime/test/providers/cpu/model_tests.cc
@@ -39,6 +39,8 @@
 #include "core/providers/armnn/armnn_provider_factory.h"
 #endif
 
+#include "test/common/cuda_op_test_utils.h"
+
 // test infrastructure
 #include "test/onnx/testenv.h"
 #include "test/onnx/TestCase.h"
@@ -94,6 +96,21 @@ TEST_P(ModelTest, Run) {
 
   std::unique_ptr<OnnxModelInfo> model_info = std::make_unique<OnnxModelInfo>(model_path.c_str());
 
+#if defined(__linux__)
+  // ORT enables TF32 in GEMM for A100. TF32 will cause precsion loss and fail this test.
+  if (HasCudaEnvironment(800) && provider_name == "cuda") {
+    per_sample_tolerance = 1e-1;
+    if (model_path.find(ORT_TSTR("SSD")) > 0 ||
+        model_path.find(ORT_TSTR("ssd")) > 0 ||
+        model_path.find(ORT_TSTR("yolov3")) > 0 ||
+        model_path.find(ORT_TSTR("mask_rcnn")) > 0 ||
+        model_path.find(ORT_TSTR("FNS")) > 0) {
+      SkipTest("Skipping SSD test for big tolearance failure or other errors");
+      return;
+    }
+  }
+#endif
+
   if (model_info->HasDomain(ONNX_NAMESPACE::AI_ONNX_TRAINING_DOMAIN) ||
       model_info->HasDomain(ONNX_NAMESPACE::AI_ONNX_PREVIEW_TRAINING_DOMAIN)) {
     SkipTest("it has the training domain. No pipeline should need to run these tests.");
diff --git a/onnxruntime/test/providers/cuda/nhwc/conv_transpose_test.cc b/onnxruntime/test/providers/cuda/nhwc/conv_transpose_test.cc
index 06da2a5304..6514feadf0 100644
--- a/onnxruntime/test/providers/cuda/nhwc/conv_transpose_test.cc
+++ b/onnxruntime/test/providers/cuda/nhwc/conv_transpose_test.cc
@@ -70,7 +70,11 @@ TYPED_TEST(CudaNhwcTypedTest, ConvTransposeNhwcBias) {
   auto op =
       ConvTransposeOp<TypeParam>{.input_dims = {1, 8, 80, 80}, .kernel_shape = {5, 5}, .channels = 16, .bias = true};
 
-  MAKE_PROVIDERS_EPS_TYPE(TypeParam)
+  if (HasCudaEnvironment(800)) {
+    MAKE_PROVIDERS_EPS(1e-2)
+  } else {
+    MAKE_PROVIDERS_EPS_TYPE(TypeParam)
+  }
 }
 
 TYPED_TEST(CudaNhwcTypedTest, ConvTransposeNhwcPad) {
diff --git a/tools/ci_build/github/azure-pipelines/linux-gpu-ci-pipeline.yml b/tools/ci_build/github/azure-pipelines/linux-gpu-ci-pipeline.yml
index 1060a0138e..5779b1da3f 100644
--- a/tools/ci_build/github/azure-pipelines/linux-gpu-ci-pipeline.yml
+++ b/tools/ci_build/github/azure-pipelines/linux-gpu-ci-pipeline.yml
@@ -137,7 +137,7 @@ jobs:
             --enable_cuda_profiling --enable_cuda_nhwc_ops \
             --enable_pybind --build_java \
             --use_cache \
-            --cmake_extra_defines  CMAKE_CUDA_ARCHITECTURES=75; \
+            --cmake_extra_defines  CMAKE_CUDA_ARCHITECTURES=86; \
               ccache -sv; \
               ccache -z"
     workingDirectory: $(Build.SourcesDirectory)
@@ -166,7 +166,7 @@ jobs:
     skipComponentGovernanceDetection: true
   workspace:
     clean: all
-  pool: Onnxruntime-Linux-GPU-T4
+  pool: onnxruntime-Linux-GPU-A10
   dependsOn:
   - Linux_Build
   steps: