DDR5 tuning - tRFC & turn around timings

Cenuij · 21 Nov 2024 at 11:32

Am still very much learning about DDR5 as this is my first AM5 board.

I came across this guide to tRFC caps on another board, but there was no reference to the source, does anyone recognise where it came from?

It was actually very helpful because on my dual rank(or is it quad rank with 2 banks per side on DDR5?) 2x 32G kit (Hynix M-die), I was seeing stress test errors with tRFC = 480 with TM5, but the same tRFC and even tighter timings (CL 28) passed an 8 hour Karhu run. tRFC = 496 passes both after a decent run

Went back to CL30, because despite an intake fan directly above the RAM CL28 needed 1.52v to pass some tests, and with max tREFI the sticks were getting a little bit toasty 58-61c. It's fine though with 64G+ expected to have to make compromises and very happy with 6200/2200.

I'm also wondering if it's worth trying to tune the turn around timings tRDRDSCL/tWRWRSCL. I've read up on what they do but I don't know if I should try and tweak them or what latency impact that might have, or what values they _should_ have outside of defaults when tuning DDR5.

Any suggestions here welcome.

WARburton · 21 Nov 2024 at 22:40

Cenuij said:
I was seeing stress test errors with tRFC = 480 with TM5,

this is because TM5 works

what ram kit do you have and what speed/timing did you actually get stable?

Cenuij · Friday at 01:34

WARburton said:
this is because TM5 works

what ram kit do you have and what speed/timing did you actually get stable?

Ah sorry if that wasn't clear from the original post, the timings above are looking stable so far with shortish tests.

Kit is trident z5, F5-6000J3040G32GX2-TZ5NR

G.Skill Trident Z5 Neo EXPO RGB 64GB (2x32GB) DDR5 PC5-48000C30 6000MHz Dual Channel Kit

Order G.Skill Trident Z5 Neo EXPO RGB 64GB (2x32GB) DDR5 PC5-48000C30 6000MHz Dual Channel Kit now online and benefit from fast delivery.

www.overclockers.co.uk

1hour TM5 (1usmus profile, previous instabilities would error after 10-25 mins or sooner), 4 hour y-cruncher all tests, 1 hour OCCT cpu+mem large dataset & AVX512

So what's interesting is CL28 (which needed 1.52v) passed an 8 hour karhu run (cache enabled), failed on TM5 within 15 mins.

Changes I've made since are vSOC from 1.48v to 1.46v, planning on a 1 hour TM5 tomorrow followed by a longer y-cruncher and I may well call it a day at that, have a CPU to OC still...

WARburton · Friday at 08:11

Yep TM5 is very good at finding errors fast. If it fails it’s not stable.

Your dumping a lot of voltage in for not much real world gain. 32gb sticks are funny in some cases you would be better with 4 x 16gb, or 2 x 24gb sticks

ATIorNvidia · Sunday at 03:50

The two sets of timings that are at 11, you need to tune them as you have dual rank sticks (they are ignored for single rank) 6-6 6-6 for both read and write should work but if not 6-6 8-8 should work.

RDRDSCL should go down to 5 or 6, WRWRSCL doesn't do anything afaik, it works at 1 and performance doesn't change.

There's nothing wrong with Karhu, one stress test is never going to give you the full picture of stability and you should continue to run multiple stress tests to verify stability (as you are doing).I just got done tuning my setup and I was failing Karhu within minutes on some configurations that were passing TM5 absolut for a few hours. Because you have 64GB though it takes a long time to get enough coverage to call it stable.

You're wasting a lot of time running the full suite of ycruncher tests, just hit it with VT3.

You should get your RAM temps under control (<55c) because it's going to be hard to determine if you're getting a heat related error (this can take hours) or if you have a timing set incorrectly.

Edit: BTW the source for trfc "calculator" is this hardwareluxx thead https://www.hardwareluxx.de/community/threads/ryzen-ddr5-ram-oc-thread.1324121/

It's quite an old guide though and most information there is somewhat outdated (it follows JEDEC spec strictly which isn't necessary) For example your TRFC or tREFI doesn't need to be multiples of anything, and you'll always get positive scaling (in terms of performance) with tighter values. trfc2 and tfrcsb isn't even used on the AM5 memory controller.