DDR5-6000 Memory Performance On Linux, Scaling From DDR5 3000 to 6000 MT/s

[V1tol]

I wonder when AMD and Intel will realize that with current amount of cores it is stupid to make only 2 memory channels? Make at least 3, or better 4. People can buy low-capacity slow memory and gain performance and capacity at an acceptable price.

[TNZfr]

It's a good summary ... but multiple memory channel are available on server processors ... it's just not the same price

[brucethemoose]

I wonder when AMD and Intel will realize that with current amount of cores it is stupid to make only 2 memory channels? Make at least 3, or better 4. People can buy low-capacity slow memory and gain performance and capacity at an acceptable price.

I bet the gains are less than you'd think in most apps, and that lots of the scaling you see here comes from reduced access times (which more channels don't get you). You're also paying far more the extra pins on the socket, traces in the mobo, and extra space on the die than you would save from using slower memory.

[andreano]

Bigger is better especially with today's high-core count designs.
The sweet spot just results from the memory controller (or wallet) not keeping up

I'm not asking about that kind of sweet spot – a point of diminishing returns due to maxing out something that's not keeping up. That was my expectation. I'm asking about datapoints that don't fit the curve: How on earth was the next fastest memory always faster than the fastest memory in every compilation benchmark?

[brucethemoose]

I'm not asking about that kind of sweet spot – a point of diminishing returns due to maxing out something that's not keeping up. That was my expectation. I'm asking about datapoints that don't fit the curve: How on earth was the next fastest memory always faster than the fastest memory in every compilation benchmark?

Probably some platform quirk. Maybe the IMC speed some multiple of the memory speed at the 2nd highest setting, or maybe some BIOS setting is automatically switched at the highest speed?


With all due respect, OC tuning in the BIOS is not Phoronix's area of expertise... not that I know anything about it either.

[mppix]

I'm not asking about that kind of sweet spot – a point of diminishing returns due to maxing out something that's not keeping up. That was my expectation. I'm asking about datapoints that don't fit the curve: How on earth was the next fastest memory always faster than the fastest memory in every compilation benchmark?

.. memory controller not keeping up ..?

Assuming this is XMP, my very wild guess would be that the bios forcing a configuration to prevent an increase of error rates..

[Shevchen]

.. memory controller not keeping up ..?

Assuming this is XMP, my very wild guess would be that the bios forcing a configuration to prevent an increase of error rates..

On boot-up, there is a standard memory training. If a setting doesn't POST, the next config is tried out. If you OC, you either POST with your OC settings or get a (very slow) fallback config so that you get back into the BIOS again to try another setting. On the verge of stability, you might POST, but get errors.

Memory OC is its own rabbit hole and I mostly do it on Windows due to a lot of OC tools being available to test stability. Once its stable, I use the setting on Linux. There are entire forums on how to properly OC memory so I leave it at that.

Long story short: On Ryzen you can boost gaming performance (esp. 1% low performance) by tweaking the timings. While most benchmarks go on average framerate, 1% low performance is imho more important. Stuttering kills the experience, even if your average framerate is "good". The default "gucci" setting on DDR4 is 3200 C14 with tuned subtimings with the help of the Ryzen DRAM-calculator. I'm using 3800-C14 with tuned subs, but my previous commenters are correct: You quickly hit a point of diminishing returns, where the price of such RAM doesn't justify the performance gained in "just gaming workloads". If you do professional work, this *might* be worthwhile, but there you want to have stability - so a bleeding edge RAM setting is a big no-no there. As of such, its just a gimmick for enthusiasts, who want to see how far technology can be pushed. (and have the pocket money for such a hobby)