Something new to tinker with: github.com/xmrig/xmrig/releases/tag/v5.2.0

1GB pages, HW prefetcher control on Intel

but only on Linux

wait, does it not require bios changing of prefetcher behaviour?

No

it's command line switch now

It changes the same MSR register as BIOS does

nice

1GB pages give +1-3% depending on CPU

+1% on Ryzen/Skylake

+3% on AMD Opterons

so running as root and adding --randomx-wrmsr --randomx-1gb-pages should enable this stuff?

or rather, don't need to run as root as long as huge pages have been configured ...

It probably changes the same since there is no proof.

hm. * 1GB PAGES disabled even with --randomx-1gb-pages

1GB pages are only available on Linux

and you need to configure them

not all CPUs support them

this is a configuration change of how Huge PAges are set up then I presume?

The easiest way to enable 1GB pages on single CPU systems: GRUB_CMDLINE_LINUX_DEFAULT="default_hugepagesz=2M hugepagesz=1G hugepages=3"

and then sudo update-grub and reboot

so checking /proc/meminfo after this, it is correct in still reporting 2M size?

but if I run as root, it will configure a few 1GB pages

check number of 1gb pages with `cat /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages`

or run "hugeadm --pool-list"

There is no "hugeadm" here.

sudo apt install hugepages

or hugepage

I don't remember exact package name

yeah found it

c# cat /sys/kernel/mm/hugepages/hugepages-1048576kB/nr_hugepages

7

(dual cpu so I sent 6 there instead of 3)

And the correct path for NUMA is /sys/devices/system/node/node0/hugepages/hugepages-1048576kB/nr_hugepages

node0, node1 and so on

and node1 for a dual socket ... right

xmrig writes to this file to allocate 1GB pages, it should work under root

hmm. still says disabled

Waiting on an apt upgrade to complete so I can install hugepages package

What CPU is it?

And how exactly do you try to enable it?

Dual E5-2670 (v1)

it should support 1GB pages

modified /etc/default/grub and ran update-grub and rebooted

Xeon E3 series don't have support

then ran sudo bash -c "echo 3 > /sys/devices/system/node/node0/hugepages/hugepages-1048576kB/nr_hugepages"

you also need to modify config.json

and sudo bash -c "echo 3 > /sys/devices/system/node/node1/hugepages/hugepages-1048576kB/nr_hugepages"

find "1gb-pages" there and set it to true

appended these to commandline: --randomx-wrmsr --randomx-1gb-pages

not using config gile

file*

# hugeadm --pool-list

Size Minimum Current Maximum Default

2097152 2560 2560 2560 *

1073741824 6 6 6

can you paste full miner console output?

kk

commandline: ./xmrig -o pool.supportxmr.com:5555 -u wallet -p xyz:bvla⊙nn --coin monero --randomx-wrmsr --randomx-1gb-pages

well, command line doesn't work apparently :D

try with config.json

hehe ok

take default one from src folder and enable it there

hm. these cpu's are L2 cache limited it seems. unexpected

I think you forgot to add a number after wrmsr in the command line

try ./xmrig -o pool.supportxmr.com:5555 -u wallet -p xyz:bvla⊙nn --coin monero --randomx-wrmsr 6 --randomx-1gb-pages

yep, confirmed

you forgot to add 6 after --randomx-wrmsr

a) it works with config.json

b) adding the 6 also enables it from command line

nice

so what's the hashrate difference?

it was not clear for me that I needed the 6 by looking at the releases here: github.com/xmrig/xmrig/releases/tag/v5.2.0

currently it looks like 0.5%

6664.8 vs 6630

note that I already did prefetcher changes in BIOS

Sandy Bridge Xeons should've got the most speedup from it. Maybe they're already limited by CPU and memory access is fast enough.

now running without root, and it says [2019-12-11 10:20:03.610] rx cannot set MSR 0x01a4 to 0x0006

not sure if it warns if those settings were already made in BIOS when run as non-root

hr is identical

xmrig doesn't revert msr value

so it's 6 from previous run

it still warns however

It warns since read/write msr requires root and xmrig can't update or read msr.

ah it can't read it either ...

Someone actually tried to mine with GPUs: 6xRX 580 - 2800 h/s, 400 W youtu.be/DnLE48xi2k8?t=50

Vega cards mining as we type: [2019-12-11 12:11:48.415] speed 10s/60s/15m 8811.7 8824.8 8828.7 H/s max 9318.2 H/s

<sech1> reddit.com/r/MoneroMining/comments/…zen_7_3700x_xmrig_520_1gb_pages_msr

I wonder if the 3950x can get higher with the msr and 1gb page tweaks:

<sech1> Yes it will. Prefetchers read more stuff from memory than needed so turning them off always improves hashrate.

sech1, What was the default value of those registers?

<sech1> I didn't check

<sech1> I can reboot now and check

How long have you bruteforced those registers to find 0x510000 and 0x1808cc16?

<sech1> I didn't, they're from l1/l2 prefetcher options on some AMD BIOS

<sech1> default values: MSR 0xC001102B = 0x2008CC17

<sech1> MSR 0xC0011022=0xC00...0002500000

<sech1> so it changes 2008CC17 -> 1008CC16

<sech1> and 0x500000 -> 0x510000

<sech1> xnbya found these options in BIOS and just took register values from there

xnbya: Have you looked into firmware asm or just advanced options in UI?

#define MSR_DC_CFG 0xC0011022

so ..22 register is data cache config

#define MSR_BU_CFG3 0xC001102B

no idea how to decipher this abbreviation

but this is for older CPUs (Bulldozer etc.), probably Ryzen just has the same registers

Found it all here: amd.com/system/files/TechDocs/42300_15h_Mod_10h-1Fh_BKDG.pdf

MSRC001_1022 Data Cache Configuration (DC_CFG)

MSRC001_102B Combined Unit Configuration 3 (CU_CFG3)

1022, bit 13, set to11=Disable the DC hardware prefetcher.

1022, bit 4 is also interesting: 1=Disable speculative TLB reloads

ohhh my gawwd, bits 10, 22:19 and 23 are even more interesting

so much to test this evening

102B register - no idea, the changed bits (28:29) are not documented there

hmm, bit numbers that are changed are different from what's written here, so I guess Ryzen changed them all

But 0xC0011023 register is also interesting

0xc001102b: 0x2008cc17 -> 0x1008cc16 or 0x2008cc17 -> 0x1808cc16 ?

0x1808cc16 at your photo and 0x1008cc16 here in irc are different values

0x1808cc16

and default ?

2008CC17

more bits has changed then

bit 28 was set to 0, bits 26:27 were set to 1

and bit 0 was set to 0

no, bit 29 was set to 0, I'm bad at counting

type binary of both values and check carefully

and bits 27:28 were set to 1

You can bruteforce only changed bits of 0xc001102b to remove any redundant bits. It's likely only 2 of them shoud be changed.

Pasting the link here: community.hwbot.org/topic/190114-ro…ii-hero-formula-impact-ln2-oc-guide

Zen Perfboost : bit.ly/2kBs15c (After loading CineBench R15 or Geekbench3 bench then run the boost, just click defaults button after benching to prevent bsod)

it also messes with MSR registers

anyone knows HR for a 2950x ?

im mining with gpus sech1 . cause im an idiot

not really, they're nice space heaters in this weather

18% hashrate increase

with the 1 gb pages?

nice

no, it's the magic MSR mod

turning off HW prefetchers

and 1GB pages, yes

yeah, memory starved systems will see a bigger jump in hashrate

is it equivalent to disabling in BIOS ?

yes

thanks

but my BIOS doesn't have this option

oh ok I see :)

hmm, I'm getting 9614 h/s now (up from 9526 h/s)

just changed "Performance Bias" option on BIOS

interesting

I guess it's some other MSR register

how to dump all MSR registers?

apply rdmsr to all values within [0xc0011000, 0xc00110ff] or even wider range, it should fail for unsupported addresses

Since there are no docs about available registers

yep

I'll try all performance bias options and then dump the best one

"aida/geekbench" performance bias got me down to 7000 h/s :D

so 9614 h/s was the best, I'll dump registers and start comparing

Do you have direct link to the firmware you're using?

cohcho dlcdnets.asus.com/pub/ASUS/mb/Socke…B450M-K/PRIME-B450M-K-ASUS-2006.zip

"Performance bias = CBR15 gentle" was the best option

I think I've found it

another MSR register, I just need to retest it to confirm

sudo wrmsr -a 0xC0011021 0x40

9601 h/s

so I think some another register adds 13 h/s

sudo wrmsr -a 0xC0011020 0

this is it

9614 h/s

MSRC001_1020 Load-Store Configuration (LS_CFG)

bit 28: 1=Disable streaming store functionality.

MSRC001_1021 Instruction Cache Configuration

looking for that last +0.1%

:D

hey, it's +0.9%

It isn't clear yet whether this is the last +0.1%

this can give more ideas: amd.com/system/files/TechDocs/42300_15h_Mod_10h-1Fh_BKDG.pdf

It's not for Ryzens, but register names there are what to look for

MSRC001_1xxx group

MSRC001_102D Load-Store Configuration 2

ForceSmcCheckFlwStDis 0=Force a

self modifying code check when a cache probe hits a store that has not retired

interesting

but probably won't give anything

I've become a bit scary when saw all those configurable things inside cpu 1 month. Are you really sure that you current cpu configuration is within 10% from the best?

1 month ago*

I'm sure that 9850 h/s is the limit for my 3700X @ 4.1 GHz

because this is where it's bottlenecked by instrution execution, not memory

I tested it at 2.05 GHz and got 4925 h/s, so this is how I calculated it

so it's within 2.5% from the best

This approach test&dump is much faster than reading firmware asm.

damn, so much progress since even 1 month ago

1 month ago 8700 h/s was the limit for me

I'm getting consistent 600+ h/s per thread, beautiful :)

1200+ h/s per physical core on all cores

getting closer to the ideal CPU ASIC performance

yeap

this is the idea, right?

I guess I can set it to 3.6-3.8 GHz, undervolt and it'll not be bottlenecked by memory anymore

nice

I wonder what 3950X could do with all this tuning...

9 bajillion h/s

18-19 kh/s is possible

yes, that's the goal, to squeeze the last bit of performance out of the CPU hardware

to make it harder for ASICs

9614 h/s, 156 W at the wall (but it's 51W when idle)

not very efficient, but I'll try to find more efficient speed/voltage tomorrow

energy-wise, the bitcoin network is roughly 360x more secure than Monero

based on the most efficient hardware

another trash article about RandomX: publish0x.com/miner-insights/cpu-mi…nates-monero-randomx-upgrade-xmrgdz

... botnets ... botnet mining ... more botnets ... battle against botnets ...