DDR performance drop with specific address access pattern

Question

Hi, 
we are seeing performance drop when we access specific address pattern from DDR. Please find the below details.

in a single test we read same address repeatedly as mentioned below

rd_addr is 64B aligned and each read transaction is rd_len=1 (128B access)

'P' is the expected performance number

rd_addr =
 
0   -> no drop (P)

2   ->  no drop (P)

4   ->  **drop (~P/4)**

6   ->  **drop (~P/4)**

8   ->  no drop (P)

10 ->  no drop (P)

12 ->  **drop (~P/4)**

14 ->  **drop (~P/4)**

..........this pattern repeats........

following  are the results if we read more than 1 address repeatedly

{0, 2} -> no drop (P)

{4, 6} -> **drop (~P/2)**

{2, 4} -> **drop (~P/2)**

{0, 2, 4, 6} -> no drop (P)

{8 ,10} -> no drop (P)

{4 ,  12} -> **drop (~P/2)**

{4, 6, 8, 10} -> no drop (P)

{4 ,6 ,12, 14} -> no drop (P)

is there any known DDR performance limitation for the above address pattern?

Thanks

Indra

Answer

Hello,

The DDR Controllers in F1 are configured to maximize the performance for incremental/sequential address patterns using the "**ROW_COLUMN_BANK_INTLV**"  addressing mode and "**Force Read and Write commands to use AutoPrecharge**" option enabled in the IP. When these options are selected, the Memory Controller issues a transaction to memory with an AutoPrecharge if Column address bit A3 is set High. Please see [PG150](https://docs.xilinx.com/v/u/en-US/pg150-ultrascale-memory-ip#page=202) for details on this configuration. Apparently, Column Address bit A3 corresponds to AXI4_Address bit[8].

If the user access patterns are to the same addresses (non-incremental pattern) then it is possible that the controller auto closes the current Bank/Row after a transaction but incurs penalty of re-opening the same bank/row and hence resulting in poor performance.

Please let us know if you have any questions.

Thanks!
Chakra

DDR performance drop with specific address access pattern

Relevant content