diff --git a/ci/matrix.yaml b/ci/matrix.yaml
index 84763b5..78184e3 100644
--- a/ci/matrix.yaml
+++ b/ci/matrix.yaml
@@ -1,7 +1,8 @@
 
-cuda_prev_min: &cuda_prev_min '11.1'
-cuda_prev_max:  &cuda_prev_max  '11.8'
-cuda_curr: &cuda_curr '12.4'
+cuda_prev_min: &cuda_prev_min '11.1' # Does not support the CUPTI APIs we uses (added in 11.3)
+cuda_prev_max: &cuda_prev_max '11.8'
+cuda_curr_min: &cuda_curr_min '12.0'
+cuda_curr_max: &cuda_curr_max '12.4'
 
 # The GPUs to test on
 gpus:
@@ -57,29 +58,48 @@ oneapi: &oneapi { name: 'oneapi', version: '2023.2.0', exe: 'icpc' }
 # Configurations that will run for every PR
 pull_request:
   nvcc:
-    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc7,     std: [17]}
-    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc8,     std: [17]}
-    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc9,     std: [17]}
-    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *llvm9,    std: [17]}
-    - {cuda: *cuda_prev_min, os: 'windows2022', cpu: 'amd64', compiler: *msvc2017, std: [17]}
-    - {cuda: *cuda_prev_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc11,    std: [17], extra_build_args: '-cmake-options -DCMAKE_CUDA_ARCHITECTURES=90'}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc7,     std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc8,     std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc9,     std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc10,    std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc11,    std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc12,    std: [17], extra_build_args: '-cmake-options -DCMAKE_CUDA_ARCHITECTURES=90a'}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc12,    std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'arm64', compiler: *gcc12,    std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm9,    std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm10,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm11,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm12,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm13,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm14,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'amd64', compiler: *llvm15,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'amd64', compiler: *llvm16,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'arm64', compiler: *llvm16,   std: [17]}
-    - {cuda: *cuda_curr,     os: 'windows2022', cpu: 'amd64', compiler: *msvc2019, std: [17]}
-    - {cuda: *cuda_curr,     os: 'windows2022', cpu: 'amd64', compiler: *msvc2022, std: [17]}
-    - {cuda: *cuda_curr,     os: 'ubuntu22.04', cpu: 'amd64', compiler: *oneapi,   std: [17]}
+    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc6,     std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc7,     std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc8,     std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *gcc9,     std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *llvm9,    std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'ubuntu18.04', cpu: 'amd64', compiler: *oneapi,   std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'windows2022', cpu: 'amd64', compiler: *msvc2017, std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_min, os: 'windows2022', cpu: 'amd64', compiler: *msvc2019, std: [17], extra_build_args: '-cmake-options -DNVBench_ENABLE_CUPTI=OFF'}
+    - {cuda: *cuda_prev_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc11,    std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc7,     std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc8,     std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc9,     std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc10,    std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc11,    std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc12,    std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu22.04', cpu: 'arm64', compiler: *gcc12,    std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm9,    std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm10,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm11,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm12,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm13,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm14,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu22.04', cpu: 'amd64', compiler: *llvm15,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu22.04', cpu: 'amd64', compiler: *llvm16,   std: [17]}
+    - {cuda: *cuda_curr_min, os: 'ubuntu22.04', cpu: 'arm64', compiler: *llvm16,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc7,     std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc8,     std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc9,     std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *gcc10,    std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc11,    std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *gcc12,    std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'arm64', compiler: *gcc12,    std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm9,    std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm10,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm11,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm12,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm13,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu20.04', cpu: 'amd64', compiler: *llvm14,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *llvm15,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *llvm16,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'arm64', compiler: *llvm16,   std: [17]}
+    - {cuda: *cuda_curr_max, os: 'windows2022', cpu: 'amd64', compiler: *msvc2017, std: [17]}
+    - {cuda: *cuda_curr_max, os: 'windows2022', cpu: 'amd64', compiler: *msvc2019, std: [17]}
+    - {cuda: *cuda_curr_max, os: 'windows2022', cpu: 'amd64', compiler: *msvc2022, std: [17]}
+    - {cuda: *cuda_curr_max, os: 'ubuntu22.04', cpu: 'amd64', compiler: *oneapi,   std: [17]}