Comments

3 comments

  • Avatar
    Candido Dessanti

    Hi,

    We are still working on an ARM build of our software. Porting the software to Arm shouldn't be too hard.

    I've already done this recently using a graviton and a T4G Tesla card.

    I'll get back to you when I have some news about it, but probably the fastest way is to use a script to compile the OS version on an ARM server and then run on GH-200. 

    Candido

    0
    Comment actions Permalink
  • Avatar
    Candido Dessanti

    Hello,

    We are currently testing an experimental version of HeavyDB that utilizes the interconnection of GH-200's CPU 2 GPU, resulting in significantly improved performance. The new version allows us to run queries that can use more than double the GPU's memory capacity with only a reasonable impact on performance. However, the degree of impact depends on whether the queries are limited by bandwidth or other factors like concurrency. We have observed little to no impact on performance when a query is limited by factors other than bandwidth, even if all the data is stored outside of the GPU memory.

    This is just a quick update, I'll come back when I'll have some news about an official ARM release of the software

    I hope this helps!

    0
    Comment actions Permalink
  • Avatar
    gggg

    Hi Candido,

    That's great news considering Nvidia recently released the Blackwell architecture which contains an NVLink-c2c same as the Grace Hopper on the GH200. These servers will probably become more popular and I think software such as HeavyDB will benefit if they optimize for them. Looking forward to testing that official release! Thank you for the quick feedback!

    0
    Comment actions Permalink

Please sign in to leave a comment.