I'm not asking for hash rates for any specific card, just asking about bus width. He's already said that he can't take full advantage of Fiji's 4096-bit bus, but the 512-bit bus of Hawaii is fine...so he must know something about optimal bus width.

@Claymore

Do you think your ZEC or ETH miners will be able to take full advantage of a 2048-bit memory bus or is it too wide?

Wow Polaris is weird. When I first heard about not being able to lower vcore below 1.0V or less than VDDCI (or whatever is happening) I didn't even believe it.

Anyway, on the overclock.net PBE thread gupsterg referred to a post that seemed like it might be helpful in that regard so just thought I'd link it here.

http://www.overclock.net/t/1604567/polaris-bios-editing-rx5xx-rx4xx/#post_25839245

This part in particular


He is talking about doing a windows regedit but says the thing he is editing is an exact copy of the powerplay table in the VBIOS.

Anyhow, not sure if this is useful info or stuff your mother already knows, but figured I'd post it in case it is.

wat

BIOS editing would be the easiest way once you have all of your clocks/volts dialed in.

Maybe this one?

https://github.com/datspike/PolarisBiosEditor

or this?

https://github.com/jaschaknack/PolarisBiosEditor

I assume ATOMBIOSReader will work for generating command and data table info to make hex editing easier but don't quote me on that.

https://github.com/kizwan/ATOMBIOSReader

Yeah it's not for editing but just fixing the checksum after you've edited it with something else like a hex editor.

Someone will probably modify it eventually. The 5xx series just came out a few days ago so it could take time. I'm not sure if the original author still works on it...looks like the last update was v1.4 in July last year.

https://github.com/caa82437/PolarisBiosEditor

Someone made v1.4.1 and added Elpida and Micron mem support though but I dunno who did that.

Anyway you can always hex edit the BIOS and then load/save it in HawaiiBiosReader to fix the checksum.

Better to use hex editor. You can see that the real straps actually begin about halfway through the lines you posted and that this BIOS supports 2 different memory types.

Try here:

http://www.overclock.net/t/1605802/official-polaris-owners-club/

Well I tried Equihash and it's really good at finding unstable memory clocks! HWiNFO reported 100 GPU memory errors or so right away on both cards at 1500MHz.

This makes way more sense, because before I bumped VDDCI/AUX up to 1.025 from 1.000, I used to only be able to run at about 1375MHz on both cards without errors. That is why I was using stilt's and 1125MHz straps, because even though they would crash above 1400, the IMC isn't good enough to run that high anyway. I wasn't getting any reported memory errors when I was mining ETH last night at 1525 though (just crashes if I went too high).

At the moment, with nerdified stock 1250MHz straps, I am getting ~385 Sol/s on the second (BFR) card @ 1100/1425
On the primary (AFR) card (it is driving my monitors atm) I am getting ~345 Sol/s @ 1075/1400

I will try reflashing with nerdified 1125MHz straps and maybe I will try The Stilt's strap again too.

EDIT: Seems to be pretty good at finding unstable core clocks as well. Just got a "GPU returned incorrect data" on the card running at 1100 core. I knew that freq was unstable because I get artifacting if I actually do any gaming or firestrike or whatever at those clocks. I will also clock these down to 1075.

EDIT2: Getting ~378 S/s @ 1075/1365 on the BFR card w/ nerdified stock 1125MHz strap
          Getting ~350 S/s @ 1075/1380 on the AFR card w/ nerdified stock 1125MHz strap <---remember this one is running my monitors so would prob be higher if ran as a secondary card.
          1075 was an unstable core frequency. The AFR card also gets ~378 S/s @ 1050/1375. Driving the monitors doesn't seem to be bothering it (during general browsing at least). I am really impressed by how good of a stability test Claymore's Zcash miner is!

https://i.imgur.com/NprszQp.jpg

Now I'm getting ~379 and ~355 respectively, so obviously it fluctuates a bit. Only 1 memory error after ~20 mins so that's acceptable.

https://i.imgur.com/glWShA8.jpg

I think it's time to start trying even lower straps. Next up: Nerdified 1000MHz straps.

EDIT3: Tried nerdified 1000MHz strap on the BFR card. It was getting a lower Sol rate @ 1250MHz and crashed at 1275MHz. I went back to nerdified 1125MHz strap and it's getting 375 - 380 Sol/s @ 1075/1370
The AFR card is doing better with the nerdified 1000MHz strap. It gets errors at 1300MHz, but @ 1075/1275 it is getting 368 - 375 Sol/s.
1075 was unstable for the core. It does better with the nerdified 1125MHz strap (~378 Sol/s @ 1050/1375).
1000MHz straps are just too tight to be useful.

I've never mined cryptonight and I haven't done equihash since Nov/Dec. I might try it again some time but who knows.

BFR straps were just from putting stock straps through nerdralph's tool so thank him.

EDIT: This is where I got the 1125MHz BFR strap from in case you want other ones:

http://www.overclock.net/t/1554360/tahiti-memory-timings-patch-for-hynix-vram/#post_24933352

Yeah I just played around with different numbers using the fools tool and figured it out heh. Already did my testing.

R9 290, 1.025VDDCI/AUX, 1100/1525, Hynix H5GQ2H24AFR

stock timings (1501 - 1625MHz strap)
7771332000000000105A7B41805511112EA59906004C060122119D086C0F14206A8900A00000012 017112B3169262F15
30.83 - 30.95MH/s

nerdified stock 1500MHz timings
7771332000000000CE515A3B705510102BA218060040060022009D08640E14206A8900A00000012 0100F272D61232C14
31.00 - 31.07MH/s

nerdified stock 1375MHz timings
7771332000000000ADCD593770550F10292198050040050022EE1C08640D14205A8900A00000012 0100E242A59222A14
~31.03MH/s

nerdified stock 1250MHz timings
77713320000000008CC5583160550F0F251E17050040040022CC1C085C0B14204A8900A00000012 0100D2025511F2613
~31.03MH/s @ 1450MHz...seemed to decrease when I raised mem freq more than this

nerdified stock 1125MHz timings
55513320000000006BBD572D60550D0E229C96040020030022BB1C08530A1420BA8800A00000012 0100C1E22491D2312
30.99 - 31.07MH/s @ 1500MHz...crashed at 1525

stilt's AFR timings
77713320000000000839472A50550C0B242045040046C40022BB1C005C0B14204A8900A00000012 0120C211E51192613
30.83 - 30.90MH/s @ 1400MHz...crashes above 1400

stilt's nerdified AFR timings
77713320000000000839572A50550C0B242045040040040022BB1C005C0B14204A8900A00000012 0100C211E51192613
30.95 - 31.04MH/s @ 1400MHz....crashes above 1400





R9 290, 1.025VDDCI/AUX, 1100/1525, Hynix H5GC2H24BFR

stock 1250MHz timings (there aren't any higher straps in the BIOS)
77713320000000008CC55835805510112C9418050048C50022FF1C086C0F14205A8900A00000012 0120D242951242D15
31.00 - 31.03MH/s

nerdified stock 1125MHz timings (got these from a Tahiti BIOS with BFR...my card doesn't have this strap)
77713320000000006BBD572F70550F10289297040040040022EE1C08640D14204A8900A00000012 0100C202449212914
31.02 - 31.05 @ 1375MHz...crashes at higher mem freq

nerdified stock 1250MHz timings
77713320000000008CC55835805510112C9418050040050022FF1C086C0F14205A8900A00000012 0100D242951242D15
~31.03MH/s @ 1525MHz
~31.73MH/s @ 1125/1525
~32.43MH/s @ 1150/1525

TL;DR I gained as much as 0.2MH/s over stock by modifying Hynix timings on Hawaii. I can gain about 0.7MH/s per every +25MHz core frequency.

I'm doing some testing right now with Claymore 9.0 in Windows 7 x64.

So far what I can tell you, with Hawaii at least, the benefit from timings is negligible. It's probably because the bus is so wide anyway, they are bottlenecked by compute. I got way bigger increases by upping the core frequency. I'll report my findings after some more tests.

But I've already been using RRD = 4 for months because I've been using Stilt's strap. I am going to test both, I just want to know how to modify the output of the tool to change RRD=5 to RRD=4 so I can start testing.

Btw my 290's max mem clock is ~1500 so I probably need tighter than usual timings. The 390s can do over 1625 so they probably need RRD to be higher.

EDIT: nvm I just realized there's a windoze version of the ohgodadecode tool so I'll just check myself.

EDIT2: mmmkay that tool says that TRRD = 0 when i input 77713320000000000839472A50550C0B242045040046C40022BB1C005C0B14204A8900A00000012 0120C211E51192613

I'm guessing it only works for RX timings.

I'm just going to assume my original...uhh...assumption is correct.

Ok but how would I mix the old RRD with the new strap the tool spits out?

K yeah you make a very good point about it being more common and easier to find solutions for. I added another question to my above post.

I've been reluctant to post the link to this thread anywhere so that it (and your inboxes) doesn't get spammed with noob questions, but I did post a link to your fools' tool over on oc.net and told them credit belongs to nerdralph @ bitcointalk but asked them not to spam you. Just thought I should let you know.

@p4n

Yes the straps are always 48 bytes long. However, it is better to search for the 3-byte strings (see below quote) to find the strap you want rather than searching for 7771. They don't all necessarily start with that. It depends on the type of RAM your GPU has and which frequency range the strap is for.

Also it's possible your GPU's BIOS supports multiple RAM types. So for example, if you want to edit the timings of the 1126-1250MHz strap, and your card supports Elpida BBBG, Hynix AFR, and Hynix BFR, then you will have a strap after 48 E8 01 01, 48 E8 01 02, and 48 E8 01 03.

Make sure you edit the correct strap for the specific RAM ICs your card has (or just edit all 3 of them to be sure).

Also if you load a BIOS in HawaiiBiosReader, and then save it, it fixes the checksum automatically. No need to use a separate tool for that.


@nerdralph

Thanks for the reply! I had started to read up on distros and read something about Ubuntu being based on an unstable Debian distro and having lots of issues, so I was thinking a stable or testing distro might be better...and then installing a GUI like Gnome or something on top if I want to use one. Debian and Arch were recommended in the thing I was reading. So Ubuntu is good for getting to learn the nuts and bolts of linux?

Also, thanks for all the work you do for the community!

EDIT:

Okay so I tried putting Stilt's straps into the tool (among many others) and I noticed Stilt had RRD at 4, but the tool changed it to 5. Stilt did have other timings for FAW and 32AW (like 6) which were changed to 0, so I assume that's an improvement, but I'm wondering if I should leave RRD at 4?

For example, Here is the results for Stilt's strap for Hynix AFR


I bolded the number that I assume is RRD. I'd like to try the new strap, but with Stilt's old RRD. If I just change that 57 back to a 47 would that accomplish it?