How to determine a good ThreadblockShape in CUTLASS

Accelerated Computing CUDA CUDA Programming and Performance
#1

Why the default configuration of GEMM in CUTLASS use a ThreadblockShape of [128, 128, 8]? I know that BlockM (128) and BlockN (128) might be determined in terms of arithmetic intensity, but why BlockK is set to 8 ?

// include/cutlass/gemm/device/default_gemm_configuration.h
template <
  typename ArchTag,
  typename ElementA, 
  typename ElementB, 
  typename ElementC, 
  typename ElementAccumulator>
struct DefaultGemmConfiguration<
  arch::OpClassSimt, 
  ArchTag,
  ElementA, 
  ElementB, 
  ElementC, 
  ElementAccumulator> {
  
  static int const kAlignmentA = 1;
  static int const kAlignmentB = 1;
  using ThreadblockShape = GemmShape<128, 128, 8>;
  using WarpShape = GemmShape<32, 64, 8>;
  using InstructionShape = GemmShape<1, 1, 1>;
  static int const kStages = 2;

  using EpilogueOutputOp = epilogue::thread::LinearCombination<
    ElementC,
    1,
    ElementAccumulator,
    ElementAccumulator
  >;

  using Operator = arch::OpMultiplyAdd;
};