diff --git a/aten/src/ATen/cuda/CUDABlas.cpp b/aten/src/ATen/cuda/CUDABlas.cpp
index 46d73bc897a5..d534ec5a1785 100644
--- a/aten/src/ATen/cuda/CUDABlas.cpp
+++ b/aten/src/ATen/cuda/CUDABlas.cpp
@@ -183,22 +183,13 @@ uint32_t _getAlignment(uintptr_t address) {
 
 static size_t _parseChosenWorkspaceSize() {
   const char * val = getenv("CUBLASLT_WORKSPACE_SIZE");
-  size_t workspace_size = 1024;
 #ifdef USE_ROCM
   if (!val) {
     // accept either env var
     val = getenv("HIPBLASLT_WORKSPACE_SIZE");
   }
-#else
-  cudaDeviceProp* p = at::cuda::getDeviceProperties(c10::cuda::current_device());
-  // Keep workspace_size = 1024 for small Ampere GPUs
-  // See https://github.com/pytorch/pytorch/pull/120925#issuecomment-1977556485
-  if (p->major == 8 && p->totalGlobalMem / 1073741824 >= 24) {
-    workspace_size = 4096;
-  } else if (p->major >= 9) {
-    workspace_size = 32768;
-  }
 #endif
+  size_t workspace_size = 1024; /* default size in KiB according to #73328 */
   if (val) {
     try {
       workspace_size = std::stoi(val);