That's correct, but if you are, for example, multiplying two 1024x1024 matrices (typical for ML loads) then you need to perform roughly 1B operations on 1M numb

That's correct, but if you are, for example, multiplying two 1024x1024 matrices (typical for ML loads) then you need to perform roughly 1B operations on 1M numbers. There is no need for loading lot of data. So we can do without fast memory in this case, can't we? Hopefully the cache and registers of GPU can handle this.