Topic: Custom RAM Timings for GPU's with GDDR5 - DOWNLOAD LINKS - UPDATED - page 4. (Read 155685 times)

interesting study you guys made.

50 watts holy macaroni,never thought of a value near that!explains why no real gains on memory. That needs serious cooling. I remember quoting you on 550s/560s consuming more on memory than on core. Couldn't figure how is that (specially that memory die is bad when it comes to heat exchange.
Surely,AUX power numbers in HW monitor apps are nothing related to this.
On a second thought, you know you can buy those aluminium heatsinks,extruded on demand (custom) and machined to dimensions so they can be seated just similar to performance waterblocks, if your modification is simple and you purchase some of these, you can get them for very low prices.

Micron and micron-elpida are quite a different story, both in timings and max clock.

Also, your estimate about memory chips power consumption is way too low,

"RX 480 uses 40-50W for its 8GB"

Source:

https://www.gamersnexus.net/guides/3032-vega-56-cost-of-hbm2-and-necessity-to-use-it


So for 8GB cards almost half the consumption goes to memory chips

i do use thermal pads (these are quite good) https://www.aliexpress.com/item/ARSYLID-High-quality-100-200-1-0-2-0-4-5mm-Thermal-conductivity-3-6W-CPU/32835108024.html?spm=a2g0s.9042311.0.0.4fd54c4d3E82Yb
0.5 mm or 1 mm and a copper shim, 1 mm thick or more https://www.aliexpress.com/item/20-pcs-15mmx15mm-0-3mm-0-4mm-0-5mm-0-6mm-0-8mm-1mm-Heatsink-Copper/32826278552.html?spm=a2g0s.9042311.0.0.3dbd4c4dpaqYyI
I experiment with the thicknesses so that the stock heatsink presses the shim and the pad slightly without being thicker than the processor height (been there done that, and the GPU fan was squealing ) the more spacious the copper shims the better, however, i do this on rough estimates, that the memory die consumes 1-2 watts each (reasonable?) the shims not touching the aluminum directly is not a big problem, because copper blown by the gpu fan imho is more than sufficient to cool the memory. Most of the Aorus 580s i have memory cooled from the factory reach 2100-2200mhz (8 gb, 4gb all of them hynix) i found thermal pads being 0.5 mm and perfectly matched with the heatsink, but, as a consequence of mining these pads will dry within months (become brittle)

Do you really clock those elpidas to these clocks? lol? how do you cool them? what sort of timings for such clocks!?

Well, i'm actually modding micron now and due to lack of resources its mostly lots of trial and error and understanding what others did.
But i'm glad to share my findings,

Micron and Elpida are quite easy (not a samsung guy, non of my cards are)
In brief, you do the following, "in Leyman's terms" shorten tRC, loosening reads slightly. I do so for trcdr,actrd, trprda and dropping writes slightly (again for trcdwr,actwr here i'm speaking about going to 14 from 16 on trcdwr for example
Elpida is hilariously funny, glad i started learning with it, no rules (almost!) , so for example for most eplidas an easy solution for low frequency memory (will get hotter though/unstable) try dropping trrd to 4 on the one click strap or SRB bios editior's pimped strap and loosening reads slightly increasing tRFC substantially (thanks to Kueto again who helped with this!) it gives you more stability for negligible speed loss, clock memory to 1875-2025 however, most non cooled memory, will give errors above 1900, still will give 545 h/s -585 h/s (560, will be 1000 h/s++ for 570/80)
The biggest challenges though, are rules.

*Starting with the 1625 strap in hynix for example:
1- the 999 and the tCKE and tCKE Pulse(19,3 or 21,5 in hynix etc.) are no touch unless you change MISC-SEQ-1 and tCL, TCR accrodingly.
2- tCR of the stock strap, needs to be loosened, its tied to tCL, tRCDR,tRP, tRRD are tied to tCR (not tRP in elpida, thats why you can have a very low tCR value) so any +1/-1 to any of those, must be subsequently, done tCR value.
example,
tcrdr 23, tRRD 7,tCL 18, tRP 22 tCR=68
trcdr 24 (+1), tRRD 6 (-1), tCL 18, tRP 22 tCR= (68)

3-RASMACTRD,RASMACTWR are tied to ACTRD and ACTWR by the rule, RASMACTRD(/W) + ACTRD(/W)= tCR+1 (check any stock strap, same thing) so when you drop writes (-actwr) you should increase RASMACTWR
4- Raw Precharge time, tRP must be higher than tCL, it doesnt affect Cryptonight much so its one of the keys to attain high stable straps (by loosening it)
5-loosening tCL is the key to get a higher strap, but now that's the biggest problem, tCL in elpida isn't tied to tCR (that's why you see that one click strap, being good with a low tRC), however, CAS timings (tCL,W2R,R2W) are linked to SEQ-MISC-1
If you changed the tCL of the 1625 strap to the 1750 one without the SEQ-MISC-1 of the 1750 one, will end in BSODs or similar.
Micron one click, took, Elpida's 1500mhz strap and mixed it, so higher tCL mainly, and that works..
6-increase tRFC substantially, have minimal impact on CN speeds, but gives more stability (+25% or whatever, noticed no real rules here.)
7-RP,WRPLUSRP keep the gap bewteen them to be taken along with tCL and the SEQ-MISC-1

atm, trying to get along linux, to run that AMDMEMINFO (thanks Nerdralph!) thast the only way to decipher those MC-SEQ-1 Write recovery and many other complicated crap, i just want to know how to correct it for example, to point to tCL=20 in the awkward Hynix AJR
Last, RASMACTWR should be increased to (+1 over tRC) if you happen to do heavy mixing with high tCL,tRP, while tight write parameters noticed this in some mods, Hynix MJR in particular, looking at the bright side though, experimenting with higher RASM values is ok (no bsod)

Glad to help with any info, i have loads of PDFs i stumbled upon online. glad to share all.

Apparently is a Micron -70 variant. It runs at 1900 with zero errors. I'm trying to improve performance per clock and after I will try to find the sweet spot between 1900 and 2000.

Well, I did some "cherry-picking" with your, Branko and PBE straps and I'm getting 918H/s at 1150/1900 in V7. Well, it's greater than PBE one.

My last custom strap is there: 777000000000000022CC1C00CE615C41C0571016B30C150800408700140514207A8900A00300000 010122F3F812C3617
ACT/RASMAC/RD/WR/TRC from Ubermix/Heavyarms
tRFC from Branko
tRCRD/A and TRCDW/A and base from Matthev
tRRD from PBE.
tFAW32 increased to 4 (more stability with no negative performance issues)

MISC values from Samsung's Ubermix gives me BSOD before run miner. I'll try other values.

If its micron -60 variant (have it in GARBAGE 560D by sapphire, 14 CU even 52% asic quality, even unlocked to 16 CU asic goes to nearly 50%) But at least in that 560D its stated that its a 1500mhz memory... still, the one click strap does 500-512 h/s and maximises at around 1925-1950 with occasional errors, i did mod one of them with thermal pads +a copper shim not much expected though... sometimes the case might be factory overclock "maybe" then your 470 comes with 1750 mhz out of the box, but will never pass 1900-2000mhz

Drooling here for one of those mining crash sales on ebay, 580 8gb Gaming X PLUS (Samsung, never owned such a thing!) its for~220 USD (in australia)

Some don't clock as high as others do.  I have couple of cards that don't go above 1950 without causing lots of memory errors and eventually crashing.  
The core on these cards need higher voltage and the cooler isn't good on these cards either which makes it even worse for stability.  Both cards are from same batch so it might be just that.

I bought it from a miner, so it came with some thermalpads and a good biosmod, but not all ICs are cooled (only 5/8) because the other three arent in the cooler contact area. I ordered some heatsinks at Aliexpress but it will take a little time to arrive (3 months. Yeah, Brazil's mail is poor).

About the memory strap: The board came to me with Matthev timing. I tried many timings (heavyarms, yours, some for anorak's) and the original one (matthev) was best (at least at 1900~2000 clocks). I did some changes/tests and for now I have it:
777000000000000022CC1C00CE615C41C0571016B30CD50900400700140514207A8900A00300000 010122F3F9D2C3617

Is there a way to improve "performance per clock" for Cryptonight? (considering that it couldn't run at 2000+)

Same as mine that goes up to 2300...but to go that high, cooling should be made
in such way that metal plate covers not only GPU, but memory, too (and replacing
usually bad factory thermal paste with quality one)

However, 2100 should be easy to achieve

Also, important is to find sweet spot between voltage and temperature (voltage increase
improves stability, but only up to the point where increased heat starts to be a problem)

https://imgur.com/gxtKJVB

D9SXD. Apparently is a 7Gbps IC

Yes, my RX460 4GB has Micron and go around 2100 (but has BGA issues). I will disassemble the RX470 to see the memory ICs to confirm (or not) my suspicions and take pictures to post here.

It would be nice to know exactly which micron is that, I have 1750MHz variant on my RX560s, and
no trouble to go over 2100, and even 2300 with loose timings

Well, Samsung, Elpida, Hynix, all of them is fine.

But I bought a RX470 that is making me to lose the hair (lol). Apparently it comes with a Micron low clock ICs. It doesn't run at 2000+ even with loose timings. With reasonable timings, I can run without issues between 1900 and 1930, but I can only reach 910H/s CN-V7 and 1000H/s CN-Heavy/Haven (at 1150/1920). Isn't a powerful result, right?

Wich timings I have to reduce/increase to improve performance in CN variants (maybe Ethash, but isn't the focus), considering the clock is a hard limitation? I have a little list, but I think is far from a complete list:
ACTRD/ACTWR, RASMACTRD/WR, TRCRD/A

But I'm not getting conclusive results. Is a painful pleasure to optimise it.

Is there another timing that I'm forgotting to change?

So elpidas arent that bad after all. i opened the card to see which memory it was and put (change) thermal pads, its a GB 2GB rx 560 gaming oc. Perfect low level card.
So apparently memory is 7 Gbps elpida BABG-F70
https://ibb.co/cE0E6y
https://ibb.co/kFPVeJ
I have 560 2GB oc those are probably f60 cause all of them do 1900MHz maximum regardless with my new strap they do 545 all of them.
On the other hand sapphire 2GB rx 560 have the worst Hynix modules i came in touch with. Funnily, sapphire core is 1300 mhz , 16 CU, the hynix modules are factory overclocked to 1750. All of them give errors with the one click ajr strap at 1900 mhz, one cant even do 1860mhz with it! Compared to hynix on my 3x aorus 580s which do 2100-2175mhz without errors.
I wish to share all my cardd i camr with to debunk the silicon lottery thing. The factory die quality 6-8 gbps (saw those samsungs 9gbps as well ending with -HC22 so, 2250mhz stock ? Have anyone saw these on a polaris? Red they are on 1070  ti)
.
Hynix customization is a big PITA. Increased trfc, but realising how tRC must be precisely linkee to trcd and trp/tcl and trrd! Took me ages. Still cant find a soft spot for those sapphire 6gbps 560 abominations. 475 h/s and they also fit LESS intensity threads compared to elpida!

Best way so far.

OverdriveNTool is my personal preferred way to undervolting.

deusofpower if you want a larger guide to undervolting, I recommend this: https://mining.help/

Did it... i'm hovering between 108 - 112 watts per gpu. (Hashrate 31.3Mh/s) I'f i try to go under 0.881v it simply makes no changes. (even if, for example, i'll try to set 780mV, it makes no real difference).

Thanks to Ohgoda people, i was able with ohgodatool to understand. Apparently with 560s (all of them doesnt have that IR3567B) the voltage in PBE is the MVDD (AFAIK this isn't changeable right? or is it useless completely? left it at stock 1000mv though) meanwhile the hidden lock on the memory controller was the VDDCI in that bios (tool is super awesome and detailed showed both entires, found VDDCI at 875 by default)   now HWinfo shows vddci 925mv not sure if it helped in improving straps stability to clock more, have to experiment but surely 875 is quite low.
Went ahead and raised VDDCI to 925mv and kept my strap experiments.
So far, 583-590 H/s CNv7 Normal from an RX 560 Elpida 2GB which isn't spectacular just a 2025 MHz with the one click tight crap. (JCE gpu miner beta, no rejects, sometimes i got rejects with erratic experimental straps)
still have to improve, just knew which direction to go, it took me over 40 hrs+++ for this damn elpida, but the *** never says no lol.
Edit: memory controller utilization improved to~54% and spikes to 57-59%.

The undervolt you actually made is just the floor voltage for core.

Actual core voltage hovers between voltage values indicated on p7 core and p2/p3 mem.  Use overdriventool disable all states except p7 and p2/p3 mem (highest) and set the voltage 925mV in p7 and 900mV in p2/p3 mem.  Check actual voltages while mining you should be hovering in between these two values.

If you undervolt too much you might just crash or drivers won't honor it

Checked it on hwinfo64... you were right. it says 1.131v. So, how do i lower that on bios? There should be something i did completely wrong...