Doug at ClusterMonkey pointed us to a Nehalem memory cheat sheet put together for Dell’s technology wiki by Jeff Layton. The writeup looks like a pretty complete summary of the various options for configuring memory on a Nehalem, and what the impact of the various options is on memory performance.
[SPONSORED GUEST ARTICLE] These are the times that try the souls of IT staffs tasked with deploying AI at scale… Effectively implementing AI at scale involves a whole host of design, build, deployment, and management considerations. It requires ….
This technology guide sponsored by PNY describes how the NVIDIA® EGX™ Platform provides a single architecture to meet every data center workload need. PNY aids in GPU selection and testing allowing partners to equip customers with the best NVIDIA EGX Platform solution based on their needs.
Featured RSS Feed More News from insideAI News
Summary:
* 1 bank x 3 memory channels per socket is 1333 MHz memory (if CPU supports it)
* 2 banks x 3 memory channels per socket is 1066 MHz memory
* 3 banks x 3 memory channels per socket is 800 MHz memory
Thus, if you are going for large memory capacity, you can lose a large amount of memory bandwidth. For example, if you choose 8 GB RDIMMS and populate 2 banks x 3 memory channels, you can get 96 GB of memory with an 8.5% loss in bandwidth.
If you choose 4 GB RDIMMS and populate 3 banks x 3 memory channels, you can get 72 GB of memory with an 22% loss in bandwidth. Or you could go with 8GB RDIMMS for 144 GB of memory, but still with the 22% loss in memory bandwidth.
Since I like to purchase large memory systems at work, this turned out to be an eye opener for me. Thanks for digging this up.