Server performance related to memory speed

[sparx]

Has anyone seen any analysis related to DDR4 memory speed performance?
Say if we use a Xeon scalable gen2 and drop from 2933MT to 2400MT using the same CPU. What is the performance drop in different applications?
Are there any good tools to measure memory performance?

 

[alex_stief]

Stream is one of the best tools to measure memory bandwidth. But that probably won't tell you anything you don't already know, when comparing different transfer rates.
Other than that, the impact of memory transfer rate on performance very much depends on the application. You can get anything from virtually zero impact up to a 1:1 scaling with transfer rate.

 

[sparx]

Yeah it was kind of the answer i was afraid of.

 

[RolloZ170]

Yeah it was kind of the answer i was afraid of.

You only need a faster PC if your neighbor has a faster one.

 

[BoredSysadmin]

this may give you a clue:

The Intel 12th Gen Core i9-12900K Review: Hybrid Performance Brings Hybrid Complexity

www.anandtech.com

 

[RolloZ170]

this may give you a clue:

the PC of your neighbor ?

 

[sparx]

Haha.. @RolloZ170 this was actually work stuff. I dont really care about memory speed for my own stuff.

 

[BoredSysadmin]

@sparx, ignoring the immature response, the link to Anandtech above shows the same CPU/application performance with different memory types.
While I understand this isn't 100% correlated to your situation, you should see the graphs could be either a draw or somewhat faster, depending on the apps. Basically what Alex said above.

 

[sparx]

@BoredSysadmin Im all about immature responses myself. Its all part of the joy of internet Yeah it all boils down to understanding the use case and testing in that environment. Its a good reference point, because it shows that DDR4 vs DDR5 has great performance benefits on certain applications. Probably to be expected.

 

[bayleyw]

Anywhere from zero to 20% loss in performance, probably averaging around single digit percentages. Take the AT results with a grain of salt, that was a huge (50% leap) in bandwidth comparing JEDEC DDR4 with JEDEC DDR5, some of those benchmarks (yChruncher, 3DPM, AIDA) are not real applications, and its really difficult as a whole to correlate that benchmark suite (or any suite of benchmarks, honestly) to real application performance.

Generally speaking, memory is so cheap that you buy the fastest memory officially supported by your machine for work, which is 2933 for Cascade Lake 1DPC and 2666 for 2DPC. Running out of memory is always worse than slow memory (lol) so if you must have 2DPC to make your application work then that's what you have to live with, but you can get hilariously large 2933 DIMMs. It's unlikely you'd run out of memory at 1DPC on a dual socket Cascade Lake system.

 

[Wasmachineman_NL]

AIDA is a terrible memory benchmark.

t. Buildzoid

and like sparx said, it all comes down to this thing: is your workload (y-cruncher for instance is extremely memory bottlenecked) highly memory intensive?

 

[sparx]

@Wasmachineman_NL I think thats one of the problems as well. I dont have full control over the workload. I liked the y-cruncher program. I think ill give that a try some day and see how that affects my hardware running in different memory speed.

 

[Wasmachineman_NL]

@Wasmachineman_NL I think thats one of the problems as well. I dont have full control over the workload. I liked the y-cruncher program. I think ill give that a try some day and see how that affects my hardware running in different memory speed.

Careful though, y-cruncher can (and WILL!) put your CPU on fire.

 

[sparx]

@Wasmachineman_NL not my hardware. I like fire. * insert menacing laughter here *

 

[bayleyw]

@Wasmachineman_NL not my hardware. I like fire. * insert menacing laughter here *

heh, sadly its also a giant in-memory FFT and basically not representative of anything except itself, so your users might get a bit cranky if you give them recommendations based on ycruncher performance