Hacker Newsnew | past | comments | ask | show | jobs | submitlogin

In general, you should be running the EXPO profile: this "upclocks" your RAM. I use the term "upclock" instead of overclock because you aren't pushing your RAM past its manufacturer's limit, you are pushing it to that limit. The JEDEC profile (the opposite of EXPO) is extremely conservative and can significantly affect performance.

This effect is compounded with Ryzen because of Infinity Fabric, which is the interconnect between the chiplets in the CPU. IF is clocked to the same speed as your RAM - so at the 6000MT/s "sweet spot" the article mentions your IF and RAM bus are running at 3000MHz. Matching IF to RAM bus is so important that increasing the RAM bus clock above the IF clock will decrease performance (unless the RAM bus clock is a multiple of the IF clock).

I haven't done the comparison on my 7950x, but on my 3950x running the IF at 1800MHz (RAM at 3600MT/s) resulted in a 10%-30% uplift in performance (depending on the workload).

So with Ryzen, if should consider purchasing RAM that is matched to the IF clock. For Ryzen 7 that is 6000MT/s. If you purchased RAM above that, you should definitely decrease the clock to match IF (you can tighten RAM timing to take advantage of the more expensive chips that you purchased, if you have the patience to).

What am I doing? I am going to disable EXPO, but I will definitely be re-enabling once this has been resolved.

Aside: What is MT/s vs MHz? DDR stands for "double data rate", meaning the MHz that everyone uses to describe the speed of RAM is a misnomer. "6000MHz" RAM actually runs at 3000MHz (but does two transfers per clock cycle). So: Mega-Transfers per Second.



>In general, you should be running the EXPO profile: this "upclocks" your RAM. I use the term "upclock" instead of overclock because you aren't pushing your RAM past its manufacturer's limit, you are pushing it to that limit. The JEDEC profile (the opposite of EXPO) is extremely conservative and can significantly affect performance.

Any source I could find says that EXPO does void warranty, as it is overclocking. https://www.amd.com/en/technologies/expo (in footnotes section at the end of page)

It might move RAM to it's manufacturer spec as you said but it does it by overclocking CPU

> So with Ryzen, if should consider purchasing RAM that is matched to the IF clock. For Ryzen 7 that is 6000MT/s. If you purchased RAM above that, you should definitely decrease the clock to match IF (you can tighten RAM timing to take advantage of the more expensive chips that you purchased, if you have the patience to).

Do you know any good benchmark (preferably Linux one) that would show those differences? I will be getting 7800X3D in few days so I want to test it for a bit.

> What am I doing? I am going to disable EXPO, but I will definitely be re-enabling once this has been resolved.

The core of the issue seems to be "just" overvolting so I'm guessing any manual overclock would still be fine. I have no need for it now (just puny 1070 to pair it with for now), but yeah, wait for new BIOS firmware seems to be best option.

> Aside: What is MT/s vs MHz? DDR stands for "double data rate", meaning the MHz that everyone uses to describe the speed of RAM is a misnomer. "6000MHz" RAM actually runs at 3000MHz (but does two transfers per clock cycle). So: Mega-Transfers per Second.

Not exactly. The specified speed is only speed of the bus between memory and CPU, not the speed of memory chip itself. DDR4 3200 and DDR5 6400 have exact same internal speed (400MHz) and this is why there isn't much of latency improvement between generations (DDR5 does have some other tricks for latency).

Also the 6400MHz is actual speed data is put on the bus, even if clock is half of that.

It's pretty much entirely so the RAM chips don't have some insane number of pins but something manageable.


> Any source I could find says that EXPO does void warranty, as it is overclocking. https://www.amd.com/en/technologies/expo (in footnotes section at the end of page)

> Overclocking and/or undervolting AMD processors and memory, including without limitation, altering clock frequencies / multipliers or memory timing / voltage, to operate outside of AMD’s published specifications will void any applicable AMD product warranty, even when enabled via AMD hardware and/or software. This may also void warranties offered by the system manufacturer or retailer. Users assume all risks and liabilities that may arise out of overclocking and/or undervolting AMD processors, including, without limitation, failure of or damage to hardware, reduced system performance and/or data loss, corruption or vulnerability. GD-106

I am not a lawyer! BUT!

AMD voiding the warranty like that sounds extremely illegal.(Depends on your jurisdiction, obviously.) EXPO is very clearly advertised on the box, documentation, and media coverage. It even comes on by default in a number of cases. That would mean the AMD Processor/Memory would have the warranty voided when you first power it on!

The EU specifically has a law covering that any advertised features are covered by a legally mandated 2 year warranty.

https://europa.eu/youreurope/business/dealing-with-customers...

Curious as to how that works.


Well, it is an advertised feature that explicitly states (althought under hidden footnote) that it breaks warranty so yeah, would be nice if someone tried to get them to court over it. I'd imagine they would be raked thru coals in EU

> It even comes on by default in a number of cases. That would mean the AMD Processor/Memory would have the warranty voided when you first power it on!

That I'd be doubtful about as AMD have no bearing about what options random motherboard vendor enables by default


I think this is something similar to that launch control sold on some cars. See https://jalopnik.com/this-is-nissans-gt-r-launch-control-war...


> Benchmark

I was still using Windows when I did everything with my 3950x. I used AIDA64 for the memory, specifically paying attention to latency (which is the strongest signal you will see for IF 1:1). Other than that it was a few tests that matched my anticipated workload: pi for stability, 3DMark for games (another decent signal), compilation speed for dev.

On Linux+7950x I just did some subjective tests (I have less time these days) with rust-analyzer responsiveness and there definitely was an improvement, but I can't quantify how much.

I might be less of an issue for a 7800 because, as far as I remember, it doesn't use chiplets.


IF maxes around ~2GHz on current chips so it is impossible to run it 1:1 on anything faster than DDR5-4000, but apparently something was improved along the way, and optimal RAM speed is at mentioned 6000 and not matching doesn't hurt as much in Zen 4

At least from this test https://www.hardwaretimes.com/ddr5-6000-memory-and-2000mhz-f...

it appears that even with mismatched clock (the 2166 IB with 6000MHz memory) latency continues to drop, and problems start if you try to run memory controller above that




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: