Forum Discussion

Altera_Forum's avatar
Altera_Forum
Icon for Honored Contributor rankHonored Contributor
14 years ago

How can i optimize the interconnect between BRAM and logic element

Hi, all

I have build a NiosII processor with SOPC tool and find that the utilization of BRAM is not efficient. My target device is EC3C25 and its BRAM has 608256Bits, But I can use about 53%. If I want to increase the

cache size of NiosII processor, QuartusII can not fit well and give error which show that project need more than 66 BRAM to fit. I think the root cause is the lack of interconnect resourse. Now the problem

in front of me is that how can i optimize the interconnect and let me take

good use of BRAM and result in good SOPC system performance.

Good guide is appreciated! Thank you for you help!

5 Replies

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    14 years ago

    to clarify, it sounds like you are trying to use a large portion of the block RAMs as an On Chip Memory for a Nios II processor. is that correct?

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    14 years ago

    Thanks thepancake for your reply, Just as your inferred, I really Build a on-chip memory by BRAM. For this case, Could you give me a good solution? Thanks again.

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    14 years ago

    the problem has to do with how Quartus fits block RAM for performance reasons, it doesn't seem to want to mux the output of the RAM blocks

    my advice is to decrease the size of the on-chip RAM until it fits or else write your own optimized HDL RAM which will include output muxes and maybe pipelines. then import the HDL as a custom component
  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    14 years ago

    note that ram %bit usage figure is very misleading. if the smallest block in your device is 9k bits then "use one single bit and the whole block is gone". ram blocks also have limitations on datawidth and if you wanted wider data then it connects several blocks together and may waste a large number of blocks. if you target optimum width you may save resource.

  • Altera_Forum's avatar
    Altera_Forum
    Icon for Honored Contributor rankHonored Contributor
    14 years ago

    that's true, but even if you count the number of RAM blocks in the device you'll run into no-fit based on the problem i described

    for example the 3c25 has 66 M9K blocks. running each in 256x32 for a Nios system (so 16,896 words), you won't be able to use all 66 blocks unless you wrap your own RAM logic