diff --git a/src/hip/HIPStream.cpp b/src/hip/HIPStream.cpp
index ec02425a..145a7ea3 100644
--- a/src/hip/HIPStream.cpp
+++ b/src/hip/HIPStream.cpp
@@ -47,13 +47,15 @@ HIPStream<T>::HIPStream(const intptr_t ARRAY_SIZE, const int device_index)
     std::cout << "Memory: DEFAULT" << std::endl;
 #endif
 
+  hipDeviceProp_t props;
+  hipGetDeviceProperties(&props, device_index);
+  check_error();
+
   array_size = ARRAY_SIZE;
-  // Round dot_num_blocks up to next multiple of (TBSIZE * dot_elements_per_lane)
-  dot_num_blocks = (array_size + (TBSIZE * dot_elements_per_lane - 1)) / (TBSIZE * dot_elements_per_lane);
+  dot_num_blocks = props.multiProcessorCount * 4;
 
   size_t array_bytes = sizeof(T);
   array_bytes *= ARRAY_SIZE;
-  size_t total_bytes = array_bytes * 3;
 
   // Allocate the host array for partial sums for dot kernels using hipHostMalloc.
   // This creates an array on the host which is visible to the device. However, it requires
@@ -63,9 +65,7 @@ HIPStream<T>::HIPStream(const intptr_t ARRAY_SIZE, const int device_index)
   check_error();
 
   // Check buffers fit on the device
-  hipDeviceProp_t props;
-  hipGetDeviceProperties(&props, 0);
-  if (props.totalGlobalMem < std::size_t{3}*ARRAY_SIZE*sizeof(T))
+  if (props.totalGlobalMem < std::size_t{3}*array_bytes)
     throw std::runtime_error("Device does not have enough memory for all 3 buffers");
 
   // Create device buffers
diff --git a/src/hip/HIPStream.h b/src/hip/HIPStream.h
index 76ef7df4..b437d2b7 100644
--- a/src/hip/HIPStream.h
+++ b/src/hip/HIPStream.h
@@ -14,27 +14,10 @@
 #include "Stream.h"
 
 #define IMPLEMENTATION_STRING "HIP"
-#define DOT_READ_DWORDS_PER_LANE 4
-
 
 template <class T>
 class HIPStream : public Stream<T>
 {
-  // Make sure that either:
-  //    DOT_READ_DWORDS_PER_LANE is less than sizeof(T), in which case we default to 1 element
-  //    or
-  //    DOT_READ_DWORDS_PER_LANE is divisible by sizeof(T)
-  static_assert((DOT_READ_DWORDS_PER_LANE * sizeof(unsigned int) < sizeof(T)) ||
-                (DOT_READ_DWORDS_PER_LANE * sizeof(unsigned int) % sizeof(T) == 0),
-                "DOT_READ_DWORDS_PER_LANE not divisible by sizeof(element_type)");
-
-  // Take into account the datatype size
-  // That is, for 4 DOT_READ_DWORDS_PER_LANE, this is 2 FP64 elements
-  // and 4 FP32 elements
-  static constexpr unsigned int dot_elements_per_lane{
-    (DOT_READ_DWORDS_PER_LANE * sizeof(unsigned int)) < sizeof(T) ? 1 : (
-     DOT_READ_DWORDS_PER_LANE * sizeof(unsigned int) / sizeof(T))};
-
   protected:
     // Size of arrays
     intptr_t array_size;