AMD Mem Tweak XL - Read/modify timings/pp/straps on the fly

[hammuh]

Thank you. This is very useful.

I checked default timing values and indeed it seems that RC = RAS + RP.
For example for Vega 64 defaults... RC(47) = RAS(33) + RP(14)... true

However I am not seeing the same for the other formula RAS = RCD + WR
For example for Vega 64 defaults... RAS(33) = RCDRD(16) + RCDWR(10)... nope.

Anyone have more information on this?

Have you managed to find a good value for RAS to go with the low RCD?

OK... I'm revealing my secrets.
Here are the formulas straight from wikipedia, not hard to find.



If you check ras, rcd, wr, rp, rc you can clearly see the formulas were used.
These values are not hocus pocus. They are calculated according to specific rules.

So if you change ras, you should change rc too and vica-versa

[1624575764]

1624575764

[1624575764]

1624575764

[Zorg33]

However I am not seeing the same for the other formula RAS = RCD + WR
For example for Vega 64 defaults... RAS(33) = RCDRD(16) + RCDWR(10)... nope.

Sure it's good.
RAS = MAX(RCDRD,RCDWR) + WR !

Anyone got a tool for displaying memory errors under linux? (other than trm)

[N2DCRYPT]

Anyone have more information on this?

I found the following web-page useful.

http://alexanderhuzar.angelfire.com/files/ram_timings.htm

HBM2 is SDRAM, so I assume similar discussions about SDRAM timings are applicable here as well. You'll note that the yellow "timing" column lists and discusses many of the parameters we're altering here.

Edit: one of the interesting things I've gathered from these tests (apologies if it's obvious to the more technically inclined) is how "subjective" pure memory frequency is. I've been quick to hate V56 stock memory, and hynix stock memory, because it's "slower", at least in terms of pure frequency. But what you notice when you look into the guts of the ram timings, is that the V56 memory timings (both Samsung and Hynix) are much tighter to begin with, which helps explain why you get higher hashrates at lower memory Mhz frequencies.

Perhaps there is little point in flashing Vega bioses anymore, now that this tool is available. Tweak the timings you have, and the "effective" data read/write speed will likely converge for most of these gpus, at least on memory bound algorithms, anyways.

[Zorg33]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150 2180h/s no matter what i try.
What is the trick there?

[heavyarms1912]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150h/s no matter what i try.
What is the trick there?

at what clocks on cnr?

[Zorg33]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150h/s no matter what i try.
What is the trick there?

at what clocks on cnr?

1100 mem, 1500 core
I just raised the core, hr went up to ~2200, but that is the max.

But maybe the miner sw is not the fastest on CN-R, it's a fork of XMRig.
Going to try TRM

Update: with teamredminer on default setting it's 2300 h/s

[clousian]

Any findings for us Polaris non bios modded folks on cryptonight people?

[heavyarms1912]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150h/s no matter what i try.
What is the trick there?

at what clocks on cnr?

1100 mem, 1500 core
I just raised the core, hr went up to ~2200, but that is the max.

But maybe the miner sw is not the fastest on CN-R, it's a fork of XMRig.
Going to try TRM

Update: with teamredminer on default setting it's 2300 h/s

okay.  was gonna mention we're most likely limited on core freq on cnr.
2500+ is possible but with very high core oc of 1.75 Mhz.  Making it lot less efficient.

[Zorg33]

okay.  was gonna mention we're most likely limited on core freq on cnr.
2500+ is possible but with very high core oc of 1.75 Mhz.  Making it lot less efficient.

Okay, i'm on the track again...
currently 2370h/s @1550MHz core and 1100MHz mem (~180W reported by driver, so sth like 220W on the wall)

[kerney666]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150h/s no matter what i try.
What is the trick there?

at what clocks on cnr?

1100 mem, 1500 core
I just raised the core, hr went up to ~2200, but that is the max.

But maybe the miner sw is not the fastest on CN-R, it's a fork of XMRig.
Going to try TRM

Update: with teamredminer on default setting it's 2300 h/s

okay.  was gonna mention we're most likely limited on core freq on cnr.
2500+ is possible but with very high core oc of 1.75 Mhz.  Making it lot less efficient.

The current public TRM release isn't the best for these modded timings with tightened latency. That said, it's most probably better than all other options . We're working on a version that we believe will take better advantage of the new fancy pants tools though. With a high core clk you can compensate and reach those max perf hashrates anyway, but with my current working version I'm doing 2318 h/s using 1407@900/1107@900, so those higher core clks aren't necessary. Note: this is on a Vega 64 LC, which can handle the super tight timings much better than my 8 x Vega 56 flashed 64.

Running a ton of tests on those Vega 56s flashed 64 tonight, feels like this new dimension will increase the complexity a lot. It was never trivial to truly trim those Vega rigs for efficiency and get everything rock solid stable, and with the additional degrees of freedom available now, man...

[Zorg33]

also I heard there's a tool that can bring down stable voltages by a LOT...

[clousian]

Current values on RX470/480/570 & 580. Any help what to change first?

GPU 0:  Ellesmere [Radeon RX 470/480/570/570X/580/580X/590]
CAS
  CL: 24          W2R: 17         R2R: 5          CCLD: 3         R2W: 31         NOPR: 0         NOPW: 0
RAS
  RC: 87          RRD: 7          RCDRA: 29       RCDR: 29        RCDWA: 17       RCDW: 17
MISC
  RFC: 219        TRP: 27         RP_RDA: 30      RP_WRA: 63
MISC2
  WDATATR: 0      T32AW: 8        CRCWL: 25       CRCRL: 3        FAW: 12         PA2WDATA: 0     PA2RDATA: 0
DRAM1
  RASMACTWR: 38   RASMACTRD: 31   ACTWR: 9        ACTRD: 16
DRAM2
  RAS2RAS: 119    RP: 28          WRPLUSRP: 34    BUS_TURN: 13
GPU 1:  Ellesmere [Radeon RX 470/480/570/570X/580/580X/590]
CAS
  CL: 24          W2R: 17         R2R: 5          CCLD: 3         R2W: 31         NOPR: 0         NOPW: 0
RAS
  RC: 87          RRD: 7          RCDRA: 29       RCDR: 29        RCDWA: 17       RCDW: 17
MISC
  RFC: 219        TRP: 27         RP_RDA: 30      RP_WRA: 63
MISC2
  WDATATR: 0      T32AW: 8        CRCWL: 26       CRCRL: 3        FAW: 12         PA2WDATA: 0     PA2RDATA: 0
DRAM1
  RASMACTWR: 36   RASMACTRD: 30   ACTWR: 9        ACTRD: 15
DRAM2
  RAS2RAS: 112    RP: 26          WRPLUSRP: 32    BUS_TURN: 13
GPU 2:  Ellesmere [Radeon RX 470/480/570/570X/580/580X/590]
CAS
  CL: 25          W2R: 19         R2R: 5          CCLD: 2         R2W: 29         NOPR: 0         NOPW: 0
RAS
  RC: 79          RRD: 7          RCDRA: 29       RCDR: 29        RCDWA: 23       RCDW: 23
MISC
  RFC: 219        TRP: 29         RP_RDA: 33      RP_WRA: 70
MISC2
  WDATATR: 0      T32AW: 8        CRCWL: 25       CRCRL: 3        FAW: 12         PA2WDATA: 0     PA2RDATA: 0
DRAM1
  RASMACTWR: 30   RASMACTRD: 27   ACTWR: 13       ACTRD: 16
DRAM2
  RAS2RAS: 119    RP: 28          WRPLUSRP: 38    BUS_TURN: 13
GPU 3:  Ellesmere [Radeon RX 470/480/570/570X/580/580X/590]
CAS
  CL: 19          W2R: 21         R2R: 5          CCLD: 2         R2W: 24         NOPR: 0         NOPW: 0
RAS
  RC: 83          RRD: 8          RCDRA: 27       RCDR: 27        RCDWA: 19       RCDW: 19
MISC
  RFC: 197        TRP: 27         RP_RDA: 30      RP_WRA: 62
MISC2
  WDATATR: 0      T32AW: 9        CRCWL: 25       CRCRL: 2        FAW: 14         PA2WDATA: 0     PA2RDATA: 0
DRAM1
  RASMACTWR: 35   RASMACTRD: 30   ACTWR: 10       ACTRD: 15
DRAM2
  RAS2RAS: 107    RP: 25          WRPLUSRP: 33    BUS_TURN: 12
GPU 4:  Ellesmere [Radeon RX 470/480/570/570X/580/580X/590]
CAS
  CL: 24          W2R: 17         R2R: 5          CCLD: 3         R2W: 31         NOPR: 0         NOPW: 0
RAS
  RC: 87          RRD: 7          RCDRA: 29       RCDR: 29        RCDWA: 17       RCDW: 17
MISC
  RFC: 219        TRP: 27         RP_RDA: 30      RP_WRA: 63
MISC2
  WDATATR: 0      T32AW: 8        CRCWL: 25       CRCRL: 3        FAW: 12         PA2WDATA: 0     PA2RDATA: 0
DRAM1
  RASMACTWR: 36   RASMACTRD: 29   ACTWR: 9        ACTRD: 15
DRAM2
  RAS2RAS: 111    RP: 26          WRPLUSRP: 31    BUS_TURN: 12

[Zorg33]

Current values on RX470/480/570 & 580. Any help what to change first?

get the optimized values from the public straps with a decoder tool to use with amdmemtweak

[clousian]

Current values on RX470/480/570 & 580. Any help what to change first?

get the optimized values from the public straps with a decoder tool to use with amdmemtweak

Yeah as a noob to all this modding that didn't even make sense. Are you saying to open something like overdriventool and find out what my current straps are, then somehow modify those straps via the winamdtweak? Thanks for your patie4nce.

[N2DCRYPT]

Guys, sorry if I missed this info prior, but how are you ensuring that timings are passed to the right gpus? I'd normally go off PCI bus IDs, but how are gpus ordered in this tool? Nothing is jumping out at me as a way to identify gpus and make sure looser timings are applied to gpus that throw errors in the mining software.

[heavyarms1912]

Guys, sorry if I missed this info prior, but how are you ensuring that timings are passed to the right gpus? I'd normally go off PCI bus IDs, but how are gpus ordered in this tool? Nothing is jumping out at me as a way to identify gpus and make sure looser timings are applied to gpus that throw errors in the mining software.

it's bus ordered
you can see the bus id on each gpu card properties in dev mgr.
Or alternatively just open OverdriveNTool which is already in sorted in bus order.

[kamisama233]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150h/s no matter what i try.
What is the trick there?

at what clocks on cnr?

1100 mem, 1500 core
I just raised the core, hr went up to ~2200, but that is the max.

But maybe the miner sw is not the fastest on CN-R, it's a fork of XMRig.
Going to try TRM

Update: with teamredminer on default setting it's 2300 h/s

okay.  was gonna mention we're most likely limited on core freq on cnr.
2500+ is possible but with very high core oc of 1.75 Mhz.  Making it lot less efficient.

The current public TRM release isn't the best for these modded timings with tightened latency. That said, it's most probably better than all other options . We're working on a version that we believe will take better advantage of the new fancy pants tools though. With a high core clk you can compensate and reach those max perf hashrates anyway, but with my current working version I'm doing 2318 h/s using 1407@900/1107@900, so those higher core clks aren't necessary. Note: this is on a Vega 64 LC, which can handle the super tight timings much better than my 8 x Vega 56 flashed 64.

Running a ton of tests on those Vega 56s flashed 64 tonight, feels like this new dimension will increase the complexity a lot. It was never trivial to truly trim those Vega rigs for efficiency and get everything rock solid stable, and with the additional degrees of freedom available now, man...

dear developer, thanks for the hark work for the great tool trm, so how to use this tools in vega 56 Samsung and hynix memory, what 's the settings will be for cnr on these cards, thanks

[Lunga Chung]

Current values on RX470/480/570 & 580. Any help what to change first?

get the optimized values from the public straps with a decoder tool to use with amdmemtweak

Yeah as a noob to all this modding that didn't even make sense. Are you saying to open something like overdriventool and find out what my current straps are, then somehow modify those straps via the winamdtweak? Thanks for your patie4nce.

No, he means find a recommended public BIOS which is good for your card and download it. Use SRBpolaris to open modded BIOS and decode the values. This way you will see what changes you need to dial in to AMDMT

[Miner Addict]

I changed a few things in the timings from todxx/kerney to (hopefully) improve compatibility with more air cooled Vega 64s w/ Samsung:



--rp 12 --rc 45 --rrds 3 --rrdl 3 --rcdrd 12 --rcdwr 5

This should also work for bios flashed 56 -> 64 w/ Samsung. However, I noticed hashrate drop overnight. If anyone has ideas lmk.

For Vega 56s w/ Samsung and original bios:

--rp 12 --rc 37 --rrds 3 --rrdl 3 --rcdrd 12 --rcdwr 5 --ras 25

Also based off of recommended settings from todxx/kerney, changed "rc" to work with Vega 56 w/ Samsung and original bios. Does not work with Hynix.

Edit: added --ras 25 so that RC = RAS + RP

Vega 56 Hynix, memory is limited for OC
Polor max 930, above gonna crash

Headache for tweak

[Miner Addict]

Very nice finding!

BTW I managed to reach 51.6mh, but in CN-R I can't get past 2150h/s no matter what i try.
What is the trick there?

at what clocks on cnr?

1100 mem, 1500 core
I just raised the core, hr went up to ~2200, but that is the max.

But maybe the miner sw is not the fastest on CN-R, it's a fork of XMRig.
Going to try TRM

Update: with teamredminer on default setting it's 2300 h/s

okay.  was gonna mention we're most likely limited on core freq on cnr.
2500+ is possible but with very high core oc of 1.75 Mhz.  Making it lot less efficient.

The current public TRM release isn't the best for these modded timings with tightened latency. That said, it's most probably better than all other options . We're working on a version that we believe will take better advantage of the new fancy pants tools though. With a high core clk you can compensate and reach those max perf hashrates anyway, but with my current working version I'm doing 2318 h/s using 1407@900/1107@900, so those higher core clks aren't necessary. Note: this is on a Vega 64 LC, which can handle the super tight timings much better than my 8 x Vega 56 flashed 64.

Running a ton of tests on those Vega 56s flashed 64 tonight, feels like this new dimension will increase the complexity a lot. It was never trivial to truly trim those Vega rigs for efficiency and get everything rock solid stable, and with the additional degrees of freedom available now, man...

So, Rig with 6 Vega 64 about 1320 Watts + fan, Mobo.= 1420 and sometimes fluctuates to ...
Imagine the power socket  could handle max  in Ampere ? Could triggered burn at Cable and Socket