Hacker News Reader

Comment by OptionX | Hacker News Reader

Back in the old DDR4 people used to say extremely high frequency ram (lets says above 3600MHz), that normally was O.C'd, would make your system unstable.

Any one knows if it still happens in these DDR5 days?

I'm asking because 6400MHz seems ludicrously high for a guy like me that just last week got some good old 2400MHz sticks for my home server and I'm wondering if it isn't marketing just to show high numbers.

View on HN

The speeds officially supported by a CPU's memory controller tend to lag behind the speeds officially supported by the fastest available memory modules. This gap got pretty wide toward the end of the DDR4 era: Intel's current CPUs that support both DDR4 and DDR5 only officially support DDR4 speeds up to 3200MT/s, but there's little reason to equip a desktop with slower than 3600MT/s memory given the state of the DDR4 market.

6400 MT/s is similarly beyond the fastest officially supported DDR5 speeds for CPUs. But the fine print reveals that this machine is actually using soldered LPDDR5, which is officially supported at 6400MT/s or higher by pretty much every processor or SoC that uses LPDDR5(x). Ditching the DIMM slots really does make it easier to operate the memory bus at higher speeds.

Star Labs needs to correct their site to not refer to LPDDR5 as DDR5, because they really are different types of memory and only one of them is upgradable by the end user.

The actual ram internally doesn't clock that fast, and many in fact run at the same rate as in the ddr4 days. The interface tackes the high rate down to a much lower rate but with more data flowing in parallel. But the ratio is locked, so if you overclock, the ram rate has to go up too, which is most likely what makes it unstable.

So the upshot is that if this 6400Mhz device is not overclocked, it shouldn't be unstable.

It's far worse than you make it out to be.

Max internal clocks on DDR4/5 are about 400MHz. There's a fundamental limit (rapidly diminishing return) to how fast you can drain and charge capacitors.

They try to make up for it by reading/sending entire row(s) at one time and sending quickly over the wire, but that's clearly reliant on cache line optimization of the program and doesn't do much to improve real-world latencies.

This is correct. I don't see how you think I was "making it out" to be better- nothing you added is inconsistent with what I wrote.

I don't have an answer to your primary question, but I can say that latency has gone up dramatically as speed has increased. I remember back in the original DDR days, CAS 2 was the stuff to get. Now, we see CAS latencies in the 30s and even 40(+?)!

Latency, as in time spent, has not really changed.

Memory numbers are just dumb.

CL32 at 6400MHz (3200MHz doubled) is 10 nanoseconds.

CL2 at 266MHz (133MHz doubled) is 15 nanoseconds.

Reply on news.ycombinator.com