AMD Memory Tweak - Read/Modify Timings on the fly!

FS at 27028 without any tricks with HBM memory timing changes and lower voltage and clocks.

 

Brutal

 

@OnnA try running firestrike with those settings...but just change the tref from 25000 to 43000 and see what happens.

 

^^ 25k is a little better than 43k
This timing is for Bank refresh, Yes it can be beneficial in some rare cases.
But for Gaming my 25k is best trade off (Speed vs Read/write refresh)

IMO 27k and more is good for 16GB Vega VII
You need to picture how VRAM is working in Your head then adjust.
Try CL16 with rest of my timings, then Play some games (also i have 900+ pts. in FS GPU score)
Was >28k, now i can beat 29k ! (with better CPU & PSU i can best ~30k in FS)

More about Timings:
-> http://alexanderhuzar.angelfire.com/files/ram_timings.htm

 

I see you guys are having fun

Awesome!

 

I see you have good knowledge about about memory timings, what would be your recommendation for a Radeon VII? I'm looking for both gaming and computational heavy operations but 'm a bit of a noob when it comes to memory timings and I'm not sure which values should be safe for testing.

 

I mixed some of your timings with mine. I have tessalation left at default and getting FS 28164 Remember back in Dec 2018 I was already able to beat 29k with tess off.

 

You have to Enable HBCC
Min. recommended by me is: 12GB (i have it at 16GB)

 

Edit only first 4 rows (adjust easy, test for artifacts - repeat)
tCL try 16 or 18 then adjust rest of the row by 1-2
tREF at max 35k (try 27k)
tXS at ~280-330
tRFC at ~255

Can't say much, i don't have Vega VII
Remember it has 4xHBM2, Vega 1 has 2xHBM2
This all matters when OC'ing, don't copy Vega 1 values

 

Will give it a shot, Thanks

 

Thanks for the numbers. Here's my FS score

https://www.3dmark.com/fs/19286183

 

Just tried it for the first time, with Onna's timings (#42) on my V56Pulse (with XFX 64 BIOS)...just great!

About 3% gain (maybe more, depending on the bench), with a few clicks.

Great tool OP, great profile Onna (next...mtrai).

 

@OnnA woot 30k on a Vega64 with tessalation off. @Dekaohtoura

https://www.3dmark.com/3dm/35993389

 

I think i have the worst v64 on the planet... it's rog strix vega and requires +950v to get even 1000 mhz of HBM stable. (Samsung HBM), core clock won't get stable past 1690 even with +1100mv... your timings aren't stable on me even with stock settings lol. Which driver are you using? I'm on 19.4.1

 

Raise you RCDRD to 16, RCDWR to 10 and RCAb to 47 and retest..those are the 3 most problomatic timings. Not all cards are gonna run those timings the same...and it also depends on the workload. Gaming vs compute. Compute aka mining and whatnot. can do lower timings on the first 2...but gaming is more sensitive.

Also keep in mind +950 is just stock value I run 1000mV on my HBM to get to 1150. As far as your core mV each core is different...it just may not be able to do overclock well.

I really do appreciate all onna does..but you always have to keep in mind...he tends to make over generalizing statements at time with nothing to back it up....such as with his settings here claiming it is stable for all. So take things at time with a grain of salt. He is a good guy and means well but can get over excited with things he finds. No offense is meat towards him.

 

Gonna retest. What are you using to test artifacts? Me personally run unigine valley to find things fast, then run up games as they are more sensitive for stability testing.
Btw, now is a matter of finding frequency vs timings combination like with ryzen... we need to find the sweet spot... the 3466 cl14 of ryzen on hbm2

 

Same 19.4.1 WHQL
Don't touch GPU Clock (try to find sweet spot and not exceed 1.1v)
As for HBM2 all Samsung are the same, You shouldn't have problems wit OC up to 1100MHz IMhO.
Try to go for easier GPU clocks with more on HBM2 at max of 956mV or 962mV.

Yup, but You need to remember that i have XTX LiQuiD -> Those are best tW Chips for Vega.
@mtrai Im stable so far for Gaming in those #42 CL14 settings, BFV MP, ME Catalys & RE2 working great.
Yesterday finished ME: C (1440p All Hyper, Blurrrr on Ultra) on those settings + HBM2 Tweak without much trouble.

Use some UE4 Game (Im using Vampyr)
1112MHz is max stable for UE4 (with Timing Mod, post #42)
For AC:O i can have even 1175MHz @975mV (Im not using that much for gaming anymore, i prefer 1125MHz @ 912mV to have cool & quiet GPU)
For 3Dmark 1200MHz is bench stable (some art. here and there, but nothing tragic, needs >1018mV, 1185MHz is fully stable on 993mV)
Some say that 1100MHz at 860mV is possible...

Note:
Every GPU is different, Strix is not the worst Vega
If re-pasted right then it is similar to Sapphire (You have tests on YT BTW)

 

Nope I did not forget you have a LC. I maynot have said things just perfect...but you got the gist...and people seem to forget each chip is different...and just cause it works for at that speeds and timings it may not work for them. They will have to test and test and test some more as we do. But they can also benefit from our testing and use it as a starting point. Same thing for ram and Ryzen.

/edit I wish I could find a way to get a waterblock on PowerColor Red Devil 64 but alas non exist for mine. I think with my GPU core and HBM I could get some crazy even better benchmarks.

 

Bro' Your skill & GPU are perfect IMhO
In the end of the day, we are Gamers not extreme overclockers
Im short in CPU Power + PSU (for this config)

When my ZEN2 arrives then i show some crazy 3Dmarks (30k with default tess x32*)
Just to show our ppl what Vega uArch really can do

* 3Dmark FS Default is x32 (pick custom & You'll see)

Note:
Cool & Quiet is my motto....

 

OT
Yeah cool and quiet, i agree. Thats why i prefer slightly lower clocks but quieter fans.
I dont have a liquid vega as you do, so imagine what happens to my vega at 100% load lol i can hear airplanes flying.
Best thing is to improve the 1% and AVG, so you can cap with RTSS at the stablest framerate possible and have a wonderful smooth experience with freesync. (Yes, even with freesync you still need to cap frameTIMEs with RTSS, it's better and has less input lag).

+1 protip i also use Chill on top of rtss +5 frames... example: game running at 90 fps with RTSS -> chill min-max 95 fps.
/OT

Do you really need that high of a frequency? To me it seems that after 1000 there aren't visible improvements.
Again, need to test out something like cl14 1050 HBM vs cl20 1120 HBM, timings and speed need perfect balance! For example 3200cl12 is very good and has the best latency on ryzen but 3466-3500 c14 is overall better!