Why is x16 slower than X8?

x16 Why is it better than x8 slow ？

One sentence answer ：x16 The memory particles are 2 individual bank group, and x8 The memory particles are 4 individual bank group：

We take Micron Of 8Gb Memory particles [2] give an example . Agree to provide 8G bit The capacity of , There are three kinds of Fuze：

Namely 2G x 4;1G x 8;512M x 16. Same capacity , The manufacturer provides these three configurations with different application scenarios ：

1.x4 It is mainly used to build high-capacity memory modules （ Think about why ）. But because of each rank Need to be 16 A particle , Therefore, the signal integrity requirements are high , Generally used in the server field .

2.x8 It is the largest shipment , The most common configuration , Most of the general consumer market is configured in this way .

3.x16 Because all it takes is 4 A particle , Applied to embedded or PCB Where space is limited , The consumer goods market is seen in low-end products .

x16 and x8/x4 The main difference is Bank Group It's halved ：

Let's take a look x8 Internal block diagram of ：

We see that there is 4 individual Group. and x16 Only 2 individual group：

Bank Group yes DDR4 Introduced , Helps improve performance . In different BG When reading or writing data , Can be partially concurrent , Less delay . Here we introduce two new Timings：tCCD_S and tCCD_L.CCD representative “Column to Column Delay”.S yes Short,L yes Long. Every Group Can work alone , Once complete 8n prefetch You don't have to wait for another group, So it's short delay, That is to say tCCD_S, It's usually 4.Group Inside , Wait a longer time every time , That is to say tCCD_L,tCCD_L It varies with different frequencies . How much is in the particles in our example ？

tCCD_L yes T10-T4, yes 6. Than tCCD_S High 50%.

If our data are very coincidental , It's all distributed in different group in ,Bank Group Will bring huge performance improvement . At best ,2 individual bank groups and 16n prefetch The same as the promotion of ,4 individual bank groups and 32n prefetch equally . If our data happens to be in one bank group in , The frequency is very high , The worst ,bank group No benefits . With the help of Bank interleave, Our actual situation is generally between the best and the worst .

x16 Only two. bank group, and x8 There are four bank group, This leads to a huge performance difference . Such as this netizen and Linus Result , some workload There are 20% As much as ！ Such a big performance difference , Enough to offset a generation or two CPU Performance improvements .