Workstation RAM speed vs capacity

Question

[r.d.jongh]

Hi

We are requesting a workstation in the lab for our Nanopore pipelines (metagenomics based) and we have ambitions towards doing image analysis with retraining AI models on our in-house data.

Here are the parts we have ordered so far:

• Intel Core i9-14900K
• MSI MPG Z790 EDGE WIFI
• MSI GeForce RTX 4090 Suprim X 24GB

And we wanted to get Corsair Vengeance 128GB DDR5 5600Mhz (4x32Gb), but it turns out these aren't supported by the motherboard. We were given two options by our supplier:

Using the same 4 RAM sticks (4x32GB DDR5) at 4000Mhz. This would mean we could get the computer delivered, as it is ready in this configuration. There is a possibility the computer may be able to run at 5600Mhz in a future (driver) update. OR Using 2 RAM sticks (2x48GB DDR5) at 5600 Mhz. These sticks would need to be ordered. The supplier expects this to work, but there is no guarantee.

We are wondering if any of you have any experience with the trade-off between RAM capacity and clock speed. Currently many of our mapping and processing tasks run into RAM capacity issues, but is there something to be said for less capacity with more speed in this instance?

Thanks in advance.

[Darked89]

Apart from the total RAM size what you may want to optimize is the the memory bandwidth. And that one can jump a lot by switching from the high end home consumer CPU+motherboard to combos oriented towards HPC servers. See i.e. AMD EPYC 9115.

Answer 1

[GenoMax]

The supplier expects this to work, but there is no guarantee.

Don't worry about RAM speed. It is not going to save you any appreciable amount of time in grand scheme of things on analyses. You want to get memory type/capacity that is stably supported NOW by your motherboard/CPU. You are not building a gaming PC and trying to wring the last bit of performance from it.

Answer 2

[Istvan Albert]

CPU clock speed is rarely (almost never) a bottleneck.

First and foremost, get the maximum amount of RAM. That way you can max out the number of parallel processes.

You gain most from running multiple processes in parallel rather than by running individual processes faster.

Also, get a fast hard drive; massive amounts of data are read and written.

[LChart]

Also RAID (striping or mirror+striping) has really helped i/o performance, especially for the kind of "read in bam file, manipulate reads, write out bam file" workflow steps.