Custom RAM Timings for GPU's with GDDR5 - DOWNLOAD LINKS - UPDATED

[VirosaGITS]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

Pretty nice, well done.

[1714508749]

1714508749

[1714508749]

1714508749

[1714508749]

1714508749

[kilo17]

you could link the ohgod decoder and checksum fixer i compiled, so some windows guys could use it too.
I compiled the tool too but that isn't working too well on win without some rewriting

Please post a link - i missed it 

[lpedretti]

Original Samsung 4G ( your particular GPU ) 1625
555000000000000022CC1C00CE596B44D0570F1531CB2409004AE700 [ 0B03 | 1420 ] 7A8900A003000000170F2E36922A3217

--> MC_SEQ_MISC1
 -- MR0
    WL = 3,  CL = 22,  TM = 0,  WR = 23,  BA0 = 0,  BA1 = 0,  BA2 = 0,  BA3 = 0
 -- MR1
    DS = 0,  DT = 1,  ADR = 1,  CAL = 0,  PLL = 0,  RDBI = 0,  WDBI = 0,  ABI = 0,
    RES = 0,  BA0 = 0,  BA1 = 1,  BA2 = 0,  BA3 = 0

Original Samsung 4G 1750
777000000000000022CC1C0010626C49D0571016B50BD509004AE700 [ 1405 | 1420 ] 7A8900A003000000191131399D2C3617

--> MC_SEQ_MISC1
 -- MR0
    WL = 4,  CL = 23,  TM = 0,  WR = 25,  BA0 = 0,  BA1 = 0,  BA2 = 0,  BA3 = 0
 -- MR1
    DS = 0,  DT = 1,  ADR = 1,  CAL = 0,  PLL = 0,  RDBI = 0,  WDBI = 0,  ABI = 0,
    RES = 0,  BA0 = 0,  BA1 = 1,  BA2 = 0,  BA3 = 0

I think you are off by +1 with the MR0 CAS latency.  SEQ_CAS_TIMING has CL=21(0x15) for the 1625 strap, and CL=22(0x16) for the 1750 strap.

Excuse my ignorance, but where is a 22 or 21 on 0x2014030B, or a 23 or 22 on 0x20140514 ?
22 binary is 10110, 21 binary is 10101, 23 binary is 10111, none of those patterns are in any of those two numbers... what am i missing?

The CAS latency in MR0 is just four bits (A3-A6), so it is based on a table lookup.  H5GQ1H24AFR supported all possible latencies from 5 to 20.  Micron's EDW4032BABG brief says "Programmable CAS latency: 6–27", so some latencies in the range cannot be programmed.  I don't have a Samsung datasheet for the K4G4 series (or any Samsung for that matter), so I would have to reverse-engineer the values from the straps by comparing the MR0 values to CL from SEQ_CAS_TIMING.

Timings samsung k4g80325FB

2000 777000000000000022CC1C0031F67E57F05711183FCFB60D006C070124081420FA8900A00300000 01E123A46DB354019
    tcl: 24=>25 - MISC1: 0x20140824   0010 0000 0001 0100 0000 1000 0010 0100

1750 777000000000000022CC1C00106A6D4DD0571016B90D060C006AE70014051420FA8900A00300000 01B11333DC0303A17
    tcl: 22=>23 - MISC1: 0x20140514   0010 0000 0001 0100 0000 0101 0001 0100

1625 555000000000000022CC1C00CE616C47D0570F15B48C250B006AE7000B031420FA8900A00300000 0190F2F39B22D3517
    tcl: 21=>22 - MISC1: 0x2014030B   0010 0000 0001 0100 0000 0011 0000 1011

1500 555000000000000022CC1C00AD595B41C0570E14B00B450A0068C70003011420FA8900A00300000 0170E2B34A42A3116
    tcl: 20=>21 - MISC1: 0x20140103   0010 0000 0001 0100 0000 0001 0000 0011

1375 333000000000000022CC1C008C515A3DC0570D132DCB74090048C7007A0014207A8900A00200000 0150D293197282E15
    tcl: 19=>20 - MISC1: 0x2014007A   0010 0000 0001 0100 0000 0000 0111 1010

1250 333000000000000022CC1C004A494937B0570C12294A94080046A700720E14207A8900A00000000 0130B252D89252A14
    tcl: 18=>19 - MISC1: 0x20140E72   0010 0000 0001 0100 0000 1110 0111 0010

1125 333000000000000022CC1C0029414831A0570C1125C9B3070046A6006A0C14206A8900A00000000 0110A21287B222614
    tcl: 17=>18 - MISC1: 0x20140C6A   0010 0000 0001 0100 0000 1100 0110 1010

1000 333000000000000022CC1C000839372B90570B102148D30600448600620A14206A8900A00000000 00F091D236D1F2213
    tcl: 16=>17 - MISC1: 0x20140A62   0010 0000 0001 0100 0000 1010 0110 0010

900  333000000000000022CC1C00E7B4362780570B0F9F072306002485005A091420DA8800A00000000 00E081A20621D2012
    tcl: 15=>16 - MISC1: 0x2014095A   0010 0000 0001 0100 0000 1001 0101 1010

600  333000000000000022CC1C00A520241A40570B0B97051204002264003A051420CA8800A00000000 00906121541151810
    tcl: 11=>12 - MISC1: 0x2014053A   0010 0000 0001 0100 0000 0101 0011 1010

400  333000000000000022CC1C006394121120570A091144B102002042002A021420AA8800A00000000 006040C0E2B10120F
    tcl: 9=>10  - MISC1: 0x2014022A   0010 0000 0001 0100 0000 0010 0010 1010

250  333000000000000022CC1C00628C110B10570A080EC3B00100204100220114209A8800A00000000 0040308091B0D0F0E
    tcl: 8=>9   - MISC1: 0x20140122   0010 0000 0001 0100 0000 0001 0010 0010

Hope it helps, boring part of the work done
BTW, i think this bios has a 2250 strap also, once i get to 2250mhz mem it gets back to 26 mh even after copying 1750 straps to 2000

[lpedretti]

Original Samsung 4G ( your particular GPU ) 1625
555000000000000022CC1C00CE596B44D0570F1531CB2409004AE700 [ 0B03 | 1420 ] 7A8900A003000000170F2E36922A3217

--> MC_SEQ_MISC1
 -- MR0
    WL = 3,  CL = 22,  TM = 0,  WR = 23,  BA0 = 0,  BA1 = 0,  BA2 = 0,  BA3 = 0
 -- MR1
    DS = 0,  DT = 1,  ADR = 1,  CAL = 0,  PLL = 0,  RDBI = 0,  WDBI = 0,  ABI = 0,
    RES = 0,  BA0 = 0,  BA1 = 1,  BA2 = 0,  BA3 = 0

Original Samsung 4G 1750
777000000000000022CC1C0010626C49D0571016B50BD509004AE700 [ 1405 | 1420 ] 7A8900A003000000191131399D2C3617

--> MC_SEQ_MISC1
 -- MR0
    WL = 4,  CL = 23,  TM = 0,  WR = 25,  BA0 = 0,  BA1 = 0,  BA2 = 0,  BA3 = 0
 -- MR1
    DS = 0,  DT = 1,  ADR = 1,  CAL = 0,  PLL = 0,  RDBI = 0,  WDBI = 0,  ABI = 0,
    RES = 0,  BA0 = 0,  BA1 = 1,  BA2 = 0,  BA3 = 0

I think you are off by +1 with the MR0 CAS latency.  SEQ_CAS_TIMING has CL=21(0x15) for the 1625 strap, and CL=22(0x16) for the 1750 strap.

Excuse my ignorance, but where is a 22 or 21 on 0x2014030B, or a 23 or 22 on 0x20140514 ?
22 binary is 10110, 21 binary is 10101, 23 binary is 10111, none of those patterns are in any of those two numbers... what am i missing?

The CAS latency in MR0 is just four bits (A3-A6), so it is based on a table lookup.  H5GQ1H24AFR supported all possible latencies from 5 to 20.  Micron's EDW4032BABG brief says "Programmable CAS latency: 6–27", so some latencies in the range cannot be programmed.  I don't have a Samsung datasheet for the K4G4 series (or any Samsung for that matter), so I would have to reverse-engineer the values from the straps by comparing the MR0 values to CL from SEQ_CAS_TIMING.

Timings samsung k4g80325FB

2000 777000000000000022CC1C0031F67E57F05711183FCFB60D006C070124081420FA8900A00300000 01E123A46DB354019
    tcl: 24=>25 - MISC1: 0x20140824   0010 0000 0001 0100 0000 1000 0010 0100

1750 777000000000000022CC1C00106A6D4DD0571016B90D060C006AE70014051420FA8900A00300000 01B11333DC0303A17
    tcl: 22=>23 - MISC1: 0x20140514   0010 0000 0001 0100 0000 0101 0001 0100

1625 555000000000000022CC1C00CE616C47D0570F15B48C250B006AE7000B031420FA8900A00300000 0190F2F39B22D3517
    tcl: 21=>22 - MISC1: 0x2014030B   0010 0000 0001 0100 0000 0011 0000 1011

1500 555000000000000022CC1C00AD595B41C0570E14B00B450A0068C70003011420FA8900A00300000 0170E2B34A42A3116
    tcl: 20=>21 - MISC1: 0x20140103   0010 0000 0001 0100 0000 0001 0000 0011

1375 333000000000000022CC1C008C515A3DC0570D132DCB74090048C7007A0014207A8900A00200000 0150D293197282E15
    tcl: 19=>20 - MISC1: 0x2014007A   0010 0000 0001 0100 0000 0000 0111 1010

1250 333000000000000022CC1C004A494937B0570C12294A94080046A700720E14207A8900A00000000 0130B252D89252A14
    tcl: 18=>19 - MISC1: 0x20140E72   0010 0000 0001 0100 0000 1110 0111 0010

1125 333000000000000022CC1C0029414831A0570C1125C9B3070046A6006A0C14206A8900A00000000 0110A21287B222614
    tcl: 17=>18 - MISC1: 0x20140C6A   0010 0000 0001 0100 0000 1100 0110 1010

1000 333000000000000022CC1C000839372B90570B102148D30600448600620A14206A8900A00000000 00F091D236D1F2213
    tcl: 16=>17 - MISC1: 0x20140A62   0010 0000 0001 0100 0000 1010 0110 0010

900  333000000000000022CC1C00E7B4362780570B0F9F072306002485005A091420DA8800A00000000 00E081A20621D2012
    tcl: 15=>16 - MISC1: 0x2014095A   0010 0000 0001 0100 0000 1001 0101 1010

600  333000000000000022CC1C00A520241A40570B0B97051204002264003A051420CA8800A00000000 00906121541151810
    tcl: 11=>12 - MISC1: 0x2014053A   0010 0000 0001 0100 0000 0101 0011 1010

400  333000000000000022CC1C006394121120570A091144B102002042002A021420AA8800A00000000 006040C0E2B10120F
    tcl: 9=>10  - MISC1: 0x2014022A   0010 0000 0001 0100 0000 0010 0010 1010

250  333000000000000022CC1C00628C110B10570A080EC3B00100204100220114209A8800A00000000 0040308091B0D0F0E
    tcl: 8=>9   - MISC1: 0x20140122   0010 0000 0001 0100 0000 0001 0010 0010

Hope it helps, boring part of the work done
BTW, i think this bios has a 2250 strap also, once i get to 2250mhz mem it gets back to 26 mh even after copying 1750 straps to 2000

Comparing it seems bits 25-28 of misc1 are tcl-5 up until 1375 where it gets back to zero, then the only bit that changes in misc8 is bit 29 that turns to one in 1500 and stays on afterwards up to 2000, so that may be the msb

[doktor83]

there is no 2250 strap, but nice card there if you can go over 2200mhz

[doktor83]

you could link the ohgod decoder and checksum fixer i compiled, so some windows guys could use it too.
I compiled the tool too but that isn't working too well on win without some rewriting

Please post a link - i missed it  

Here you are :


The decoder tool for windows :

https://mega.nz/#!aJAkwAxK!74I_6-f628CKIVl2TZBNrjAoujgSrKOgdnsnlwlKS_w

Checksum fixer :

https://mega.nz/#!zIwVXL4B!IUQOXRqWDw6L15rvN5gBgBN0hxc6ncQu5YcnG41m90g

[lpedretti]

there is no 2250 strap, but nice card there if you can go over 2200mhz

I'll check tomorrow, as it's so strange that the hashrate change happens exactly at that mem clock

[kilo17]

The decoder tool for windows :

https://mega.nz/#!aJAkwAxK!74I_6-f628CKIVl2TZBNrjAoujgSrKOgdnsnlwlKS_w

Checksum fixer :

https://mega.nz/#!zIwVXL4B!IUQOXRqWDw6L15rvN5gBgBN0hxc6ncQu5YcnG41m90g

Done 

[doktor83]

The decoder tool for windows :

https://mega.nz/#!aJAkwAxK!74I_6-f628CKIVl2TZBNrjAoujgSrKOgdnsnlwlKS_w

Checksum fixer :

https://mega.nz/#!zIwVXL4B!IUQOXRqWDw6L15rvN5gBgBN0hxc6ncQu5YcnG41m90g

Done 

Maybe i will rewrite the ohgodtool too if i catch some free time, to work with windouze.

Wolf, we are waiting for the mem.timings "tutorial" you and ohgodagirl promised

[bardacuda]

Woah epic thread here

I just skimmed through it since 90% is way over my head, so I might have missed it, but just wondering if any more efficient mem straps have been discovered for H5GC2H24BFR, H5GQ2H24AFR, or EDW2032BBBG that are stable in the ~1500MHz range for 290 cards. I know for the latter two The Stilt already posted modified straps for those so I'm more interested in the BFR straps...but I am curious about the other two too if Stilt's timings have since been improved upon.

Currently the stock 1250 straps for the BFR are

Code:

77 71 33 20 00 00 00 00 8C C5 58 35 80 55 10 11 2C 94 18 05 00 48 C5 00 22 FF 1C 08 6C 0F 14 20 5A 89 00 A0 00 00 01 20 12 0D 24 29 51 24 2D 15

I went ahead and changed it with a stock 1125 strap for the same ICs I found somewhere

Code:

77 71 33 20 00 00 00 00 6B BD 57 2F 70 55 0F 10 28 92 97 04 00 48 C4 00 22 EE 1C 08 64 0D 14 20 4A 89 00 A0 00 00 01 20 10 0C 20 24 49 21 29 14

All that did was decrease the max clocks so it just gets the same performance at lower clocks.

Anyway, again, just wondering if this info has already been posted and I just missed it. Keep up the great work! Is there a general donation address us plebs can send to, that is shared among everyone involved putting the time in for this?

[Truthchanter]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

This tool is working great for me so far on 2 different gpus (one samsung and one hynix). Will this most likely be stable for most gpu mem chips (new non ref cards with different hynix/samsung chips?)

[bardacuda]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

This tool is working great for me so far on 2 different gpus (one samsung and one hynix). Will this most likely be stable for most gpu mem chips (new non ref cards with different hynix/samsung chips?)

Nice! I will try that out and report back. Just waiting for some replacement fans and thermal pads to arrive first. VRMs are getting too hot on the BFR card in its current state.

EDIT: Just tried inputting Stilt's AFR and BBBG timings for shits and giggles and it did spit out something different. Does this mean it actually made them better?

Old, new RRD: 0 , 5
Old, new ACTRD: 12 , 0x10
Old, new FAW: B , 0
Old, new 32AW: 0 , 0
77713320000000000839472A50550C0B242045040046C40022BB1C005C0B14204A8900A00000012 0120C211E51192613
77713320000000000839472A50555C0B242045040046C40022B00C005C0B14204A8900A00000012 0100C211E51192613

Old, new RRD: 0 , 5
Old, new ACTRD: 11 , 0x10
Old, new FAW: A , 0
Old, new 32AW: 64 , 0
77713320000000004ABD472A60550F0F231D87030046C40022AA1C085C0B14204A8900A00000012 0110D20234A1D2411
77713320000000004ABD472A60555F0F231D87030046C40022A00C005C0B14204A8900A00000012 0100D20234A1D2411

I guess I'll find out soon enough!

EDIT2: Yeah agree with cryptominer420's post below. If I switch back to ETH or ZEC and this is gonna make the magical internet moneys flow in faster then I'd be all for sending some % of the extra income back to the people making it possible. I'm just not sure what amount should go to who. Perhaps you guys could get together and agree upon a split, and make a smart contract donation address that splits the donations accordingly.

[cryptominer420]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

Well it works, got my hynix cards up-to 28.3MH ETH and my samsung cards up-to 31.5MH Eth all 4GB powercolor cards.

Also to all the people making tools for us here it would be nice to have donation addresses so that we can send you some tips.

[nerdralph]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

This tool is working great for me so far on 2 different gpus (one samsung and one hynix). Will this most likely be stable for most gpu mem chips (new non ref cards with different hynix/samsung chips?)

Nice! I will try that out and report back. Just waiting for some replacement fans and thermal pads to arrive first. VRMs are getting too hot on the BFR card in its current state.

EDIT: Just tried inputting Stilt's AFR and BBBG timings for shits and giggles and it did spit out something different. Does this mean it actually made them better?

The tool is for Polaris straps only.  I'd have to modify it to support R9 cards.

[nerdralph]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

This tool is working great for me so far on 2 different gpus (one samsung and one hynix). Will this most likely be stable for most gpu mem chips (new non ref cards with different hynix/samsung chips?)

Wolf mentioned a card (XFX?) that doesn't like 32AW=0.  Otherwise, it should be fine if you start with a reasonable strap.  If you start with a Samsung 1500 strap and run it at 2100, it probably won't work.  1625 might if you are lucky, and 1750 will probably be solid.

[Truthchanter]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

Well it works, got my hynix cards up-to 28.3MH ETH and my samsung cards up-to 31.5MH Eth all 4GB powercolor cards.

Also to all the people making tools for us here it would be nice to have donation addresses so that we can send you some tips.

Are you using the 1500 strap for hynix and 1750 for samsung? and what mem clocks did you use to get those (specifically the samsung)?

For me I use the above straps and clock hynix at 1900-1925 (1950 if its stable) and 2050-2100 for samsung (although Im wondering if i can go higher than 2100 on some samsung chips)

[nerdralph]

Well it works, got my hynix cards up-to 28.3MH ETH and my samsung cards up-to 31.5MH Eth all 4GB powercolor cards.

Also to all the people making tools for us here it would be nice to have donation addresses so that we can send you some tips.

Are you using the 1500 strap for hynix and 1750 for samsung? and what mem clocks did you use to get those (specifically the samsung)?

For me I use the above straps and clock hynix at 1900-1925 (1950 if its stable) and 2050-2100 for samsung (although Im wondering if i can go higher than 2100 on some samsung chips)

Try the 1625 strap for Hynix.  On my Strix, 1500 was stable up to 1900, while the 1625 strap is rock solid at 2025.

[madLyfe]

Well it works, got my hynix cards up-to 28.3MH ETH and my samsung cards up-to 31.5MH Eth all 4GB powercolor cards.

Also to all the people making tools for us here it would be nice to have donation addresses so that we can send you some tips.

Are you using the 1500 strap for hynix and 1750 for samsung? and what mem clocks did you use to get those (specifically the samsung)?

For me I use the above straps and clock hynix at 1900-1925 (1950 if its stable) and 2050-2100 for samsung (although Im wondering if i can go higher than 2100 on some samsung chips)

Try the 1625 strap for Hynix.  On my Strix, 1500 was stable up to 1900, while the 1625 strap is rock solid at 2025.

when i tried this strap:

Evilop;

some love for my ref MSI RX480 4gb hynix please?

1500: 777000000000000022339D00CE516A3D9055111230CB4409004AE600740114206A8900A00200312 0150F292F94273116
777000000000000022339D00CE5153379055111230CB440900208400740114206A8900A00200312 0150F292F94273116
1625: 999000000000000022559D0010DE7B4480551312B78C450A004C0601750414206A8900A00200312 018112D34A42A3816
1750: 999000000000000022559D0031627C489055131339CDD50A004C06017D0514206A8900A00200312 019123037AD2C3A17
2000: BBB000000000000022889D0073EE8D53805515133ECF560C004E26017E0514206A8900A00200312 01C143840C5303F17

Thank you..

Small change

(just so you know, i'm not going to do this on a daily base.. :p)

Greetings!

Thank you Eliovp.
Try it not stable for me, but after change it from 79 to 98 it got stable with same speed.
777000000000000022339D00CE5153398055111230CB440900208400740114206A8900A002003120150F292F94273116
Maybe cause my card not good enough, or cause I still playing game with my pc.

777000000000000022339D00CE515A398055111230CB440900408400740114206A8900A00200312 0150F292F94273116 # small fix

i wasnt able to find stable, tough i only tested down to 1950 for mem. so maybe it would have been stable at 1900 like you said. it does 735 with no mem clock change. your stock 1625 strap ran through the CGI gets you over 2000 mem?

[nerdralph]

your stock 1625 strap ran through the CGI gets you over 2000 mem?

Yes.  Here's the stock Hynix H5GC4 1625 strap:
999000000000000022559D0010DE7B4480551312B78C450A004C0601750414206A8900A00200312 018112D34A42A3816

And here's what I'm running at 2025 (stable after >24H):
999000000000000022559D0010DE5B4480551312B74C450A00400600750414206A8900A00200312 010112D34A42A3816

[bardacuda]

For the lazy fools out there, I whipped up a cgi version of my strap mod utility.

Code:

http://doreymills.ca/cgi-bin/strapmod?555000000000000022CC1C00AD515A3EC0570E142D4A64080048C700030114207A8900A003000000150E2A3186272E16

This tool is working great for me so far on 2 different gpus (one samsung and one hynix). Will this most likely be stable for most gpu mem chips (new non ref cards with different hynix/samsung chips?)

Nice! I will try that out and report back. Just waiting for some replacement fans and thermal pads to arrive first. VRMs are getting too hot on the BFR card in its current state.

EDIT: Just tried inputting Stilt's AFR and BBBG timings for shits and giggles and it did spit out something different. Does this mean it actually made them better?

The tool is for Polaris straps only.  I'd have to modify it to support R9 cards.

Heh I thought so but thought I'd try it anyway just to see what happened. (btw nothing happened..same performance and stability as Stilt's original strap)
It's probably not worth spending time on R9 cards at this point. I'm sure most people have switched to RX cards by now for the efficiency...and anyway everything will probably be HBM before too long after Vega and Volta become the norm.