asabot
commented
2 years ago Hello,
Thank you for your questions and interest in CAKE.
For 1: CAKE performs better for the smaller and skewed matrices because they have a lower arithmetic intensity (operations per memory access). CAKE increases the arithmetic intensity of small matrices through block shaping. When the matrices are larger and less skewed the arithmetic intensity before block shaping is higher.
For 2: At the time of writing AMD uprof did not provide a counter to measure the number of DRAM accesses so we had to estimate using "number of L1 cache-line refills from DRAM". The bandwidth should be constant but the estimation method causes it to not be constant.
For 3: The bandwidth is constant because of the constant bandwidth block increasing the amount of computations associated with each core.
FuncJ
vnatesh
Hi, I think your article is very innovative and interesting. I have a few questions about the experiment. 1 Why is CAKE better than MKL at the scale of the matrix shown in the shaded area? Especially in the bottom left corner of each graph, Cake performes twice as well as MKL. Does that mean MKL's performance is less than 50% of the peak performance of the system?
2 AMD Ryzen 9 5950X is a platform with sufficient cache bandwidth. Why is CAKE's memory bandwidth requirement not constant as the number of cores increases？
3 ARM v8 Cortex A53 is a platform without sufficient cache bandwidth. Why is CAKE's memory bandwidth requirement constant as the number of cores increases？
Thanks a lot.