Zen4 voltages

[cyring]

Specifications

Known addresses

/*		--- SMU SVI [ 0x5a00c ; 0x5a010 ; 0x5a014 ; 0x6f038] ---
 *				[ CPU addr]	[ SoC addr]
 * ZEN	[8F_01h]		[ 0x5a00c ]	[ 0x5a010 ]
 * ZEN(+) [8F_08h]		[ 0x5a00c ]	[ 0x5a010 ]
 * ZEN(+) [8F_11h ; 8F_18h]	[ 0x5a00c ]	[ 0x5a010 ]
 * ZEN2 [8F_71h]		[ 0x5a010 ]	[ 0x5a00c ]
 * ZEN2 [8F_60h]		[ 0x6f038 ]	[ 0x6f03c ]
 * ZEN2 [8F_31h]		[ 0x5a014 ]	[ 0x5a010 ]
 * ZEN3 [9F_21h]		[ 0x5a010 ]	[ 0x5a00c ]
 */

/*				--- SMU SVI [ Rembrandt ] ---
 *				[ CPU addr]	[ SoC addr]
 * ZEN3(+) [AF_44]		[ 0x6f010 ]	[ 0x6f014 ]
 */

/*
 * where addr = { 0x5a04c ... 0x5a04f || 0x5a050 ... 0x5a053 }
 * and '_mod' register offset could be equaled to:
 * 0x0		: Zen & Zen+		[UNTESTED]
 * 0x1		: Zen2/Matisse		[VERIFIED]
 * 0x2		: Zen2/CastlePeak	[UNTESTED]
 * 0x5404	: Renoir		[UNTESTED]
*/

Dump

Prerequisite

• Compile zencli source code
  zencli

Scan actions

## Zen{1,2,3} Desktop, Server
for addr in 0x5a00c 0x5a010 0x5a014 ; do ./zencli smu ${addr}; done

## Ryzen Mobile
for addr in 0x6f038 0x6f03c ; do ./zencli smu ${addr}; done

## Some Core VID
for addr in 0x5a04c 0x5a050 ; do ./zencli smu ${addr}; done

Remark:
A register output value of 0xffffffff means "does not exist".

[amfern]

Hi, here is the output of the commands

## Zen{1,2,3} Desktop, Server
[root@ilya C]# for addr in 0x5a00c 0x5a010 0x5a014 ; do ./zencli smu ${addr}; done
[0x0005a00c] READ(smu) = 0x00000000 (0)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000
[0x0005a010] READ(smu) = 0x00000000 (0)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000
[0x0005a014] READ(smu) = 0x00000000 (0)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000


## Ryzen Mobile
[root@ilya C]# for addr in 0x6f038 0x6f03c ; do ./zencli smu ${addr}; done
[0x0006f038] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111
[0x0006f03c] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111


## Some Core VID
[root@ilya C]# for addr in 0x5a04c 0x5a050 ; do ./zencli smu ${addr}; done
[0x0005a04c] READ(smu) = 0x00000000 (0)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000
[0x0005a050] READ(smu) = 0x00000000 (0)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000 0000

[amfern]

The numbers, what do they mean?

[cyring]

Those zero would mean the SMU has nothing at the requested address.

0xffffffff which is the hexadecimal of decimal -1 can mean there is no register behind the address or the privilege level or reader is not sufficient

[cyring]

Unfortunately, in above dump, no meaningful values are showing up.

Can you now try to read MSR 0xC0010071
defined as COFVID

rdmsr -ax 0xC0010071

Bits of interest could be CurNbVid; presuming that Northbridge Voltage ID is somehow the SoC.

[amfern]

[root@ilya C]# rdmsr -ax 0xC0010071
rdmsr: CPU 0 cannot read MSR 0xc0010071

[cyring]

This MSR from legacy architectures doesn't work for Zen

[cyring]

In Wiki > Documentation > AMD Developer Guides you will find link to AMD Tech Docs where you can download the PPR specifications of Zen4. You have to select document of Raphael family and model numbers.
Document will list Registers of MSR or SMU

[cyring]

Processor Programming Reference (PPR) for AMD Family 19h Model 61h, Revision B1 Processors

[amfern]

Thank you, i kinda guessed the same that the 7950x CPU is from 19h family and of the 61h model because the /proc/cpuinfo lists

cpu family  : 25
model       : 97
model name  : AMD Ryzen 9 7950X 16-Core Processor

model 97 is 0x61 in hex.
But it took me quite a while to read through the PPR, i am quite new to the terminology and PPR documentations

[cyring]

Can you read those from Rembrandt ?

for addr in 0x0006f010 0x0006f014 0x0006f018 0x0006f01c ; do ./zencli smu ${addr}; done

[amfern]

root@ilya C]# for addr in 0x0006f010 0x0006f014 0x0006f018 0x0006f01c ; do ./zencli smu ${addr}; done
[0x0006f010] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111
[0x0006f014] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111
[0x0006f018] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111
[0x0006f01c] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111

[amfern]

Is it possible that the SoC voltage is reported by the Motherboard and not part of the CPU registers?

[cyring]

Usually when reported by motherboard, it goes through a Super I/O chip.
Mine is for instance a NCT

When part of Processor, at least up to the Zen3+, SoC Voltage VID goes through the SMU, retrieved either from a SMM address, either from PM table, and/or HSMP reserved to server EPYC and some HEDT

[amfern]

The motherboard is a Biostar racing GTQ b650e and i can see the super I/O chips but i am not sure how to probe them.

[root@ilya C]# superiotool
superiotool r4.19-1624-gd1c61a8e704
Found Aspeed AST2400 (id=0x00) at 0x2e
Found Aspeed AST2400 (id=0x00) at 0x4e
Found ITE IT8625E (id=0x8625, rev=0xc) at 0x2e

[cyring]

Are you already getting voltage values using the semsors command; especially SoC voltage ?

[amfern]

No, i don't see any voltage.

[root@ilya C]# sensors
drivetemp-scsi-3-0
Adapter: SCSI adapter
temp1:        +44.0°C
                                                                                                                                                                                                                                                                                                                
drivetemp-scsi-1-0
Adapter: SCSI adapter
temp1:        +46.0°C  (low  =  +0.0°C, high = +70.0°C)
                       (crit low =  +0.0°C, crit = +70.0°C)
                       (lowest = +30.0°C, highest = +50.0°C)
                                                                                                                                                                                                                                                                                                                
amdgpu-pci-0300
Adapter: PCI adapter
vddgfx:      538.00 mV
fan1:         675 RPM  (min =    0 RPM, max = 3300 RPM)
edge:         +43.0°C  (crit = +100.0°C, hyst = -273.1°C)
                       (emerg = +105.0°C)
junction:     +48.0°C  (crit = +110.0°C, hyst = -273.1°C)
                       (emerg = +115.0°C)
mem:          +62.0°C  (crit = +108.0°C, hyst = -273.1°C)
                       (emerg = +113.0°C)
PPT:          32.00 W  (cap = 290.00 W)
                                                                                                                                                                                                                                                                                                                
acpitz-acpi-0
Adapter: ACPI interface
temp1:        +64.0°C  (crit = +127.0°C)
                                                                                                                                                                                                                                                                                                                
k10temp-pci-00c3
Adapter: PCI adapter
Tctl:         +64.0°C
Tccd1:        +51.0°C
Tccd2:        +46.4°C
                                                                                                                                                                                                                                                                                                                
drivetemp-scsi-2-0
Adapter: SCSI adapter
temp1:        +50.0°C
                                                                                                                                                                                                                                                                                                                
drivetemp-scsi-0-0
Adapter: SCSI adapter
temp1:         +0.0°C  (low  =  +0.0°C, high = +70.0°C)
                       (crit low =  -5.0°C, crit = +75.0°C)
                       (lowest =  +0.0°C, highest =  +0.0°C)
                                                                                                                                                                                                                                                                                                                
nvme-pci-0c00
Adapter: PCI adapter
Composite:    +45.9°C  (low  = -273.1°C, high = +84.8°C)
                       (crit = +84.8°C)
Sensor 1:     +45.9°C  (low  = -273.1°C, high = +65261.8°C)
Sensor 2:     +47.9°C  (low  = -273.1°C, high = +65261.8°C)

[amfern]

Does CoreFreq currently able to display the SOC voltage for zen3 cpus?

[amfern]

It seems like k10temp driver used to support zen3 SoC voltage, but this support for was dropped here torvalds/linux@0a4e668
and it was added here torvalds/linux@55163a1

It seems like zen3 and zen4 are both from 19h fameliy so i guess they will have similar way of reading SoC voltage.

[cyring]

Socket, Core and SoC voltages I have achieved on Zen3+ Rembrandt Mobile
Ryzen 5 6600H

Those SMU addresses which however belong to the Mobile segment have not worked in our previous zencli tests.

This is where we are stopped with Zen4

[cyring]

And you will also read in kernel at

#define ZEN_SVI_BASE 0x0005A000

which is the same specified base address, for Desktop and Server segment, used in corefreqk.ko module.

This is where we are stopped with Zen4

If we are lucky, the manufacturer has just shifted the whole voltage BAR offset.
This is how I have, by chance, found it on Rembrandt on testing, idle, next stressed, next idle, one by one, each address, per increment of 4, because SMU is still 32-bits, at least up to Zen3+

When a SMU address has a moving value depending on system load, then you know you are facing some kind of indicator: voltage or temp or power or clock

[cyring]

And there's also a bad chance Phoenix voltages are not available either.
But I have not received some screenshot evidences of CoreFreq PHX support.

[amfern]

I have scanned the Range of addresses and I found two values that change relative to system performance. 0x000730d8 and 0x00073010

I am able to see package power via turbostat command, so I wonder if it's possible to find SoC VID in msr registers, also what is the relationship between MSR registers and PCI (what is zencli is reading)?
Screencast from 2023-05-16 16-59-54.webm
Screencast from 2023-05-16 18-24-53.webm

[amfern]

I have scanned the Range of addresses and I found two values that change relative to system performance. 0x000730d8 and 0x00073010

Anyway, how do we move forward to make sense out of those values, or maybe i should keep checking for more addresses

[cyring]

In none MSR I've found SoC voltage.
Only CPU Core VID is found within P-states and Boosted P-state MSR
SoC VID can be retrieved from undocumented SMU registers.

VID is usually an 8-bits you will supply to a voltage formula:

• Families 17h, 19h
  https://github.com/cyring/CoreFreq/blob/a28ea7bea3cab13493c1011d28fa2d564722cd3a/coretypes.h#LL726C15-L726C15
• Rembrandt
  

  CoreFreq/coretypes.h

  Line 729 in a28ea7b

  #define COMPUTE_VOLTAGE_AMD_RMB(Vcore, VID) \

The register values you found in SMU addresses 0x00073010 and 0x000730d8 you mask the 8-bits at non zero. Can be at position 0, 8, 16, or 24 then right shifted to bit position 0

Then compute Voltage using formulas (which are solving Vcore but also SoC voltage, if we're lucky)

MSR may cache PCI registers or other IP registers, can be IP microcode.
But, contrary to Intel, I have not found (many) AMD Zen overlapping features between MSR and PCI.

[cyring]

Idle

0x3bc10000

$$3B: 1.55−(0.00625×59)=1.18V$$ $$C1: 1.55−(0.00625×193)=0.34V$$

0x0000ba81

$$BA: 1.55−(0.00625×186)=0.39V$$ $$81: 1.55−(0.00625×129)=0.74V$$

Under stress

0x34010000 

$$34: 1.55−(0.00625×52)=1.22V$$ $$01: 1.55−(0.00625×01)=1.54V$$

0x0000b401

$$B4: 1.55−(0.00625×180)=0.42V$$

[cyring]

Do they look like voltages of your Processor Elements ?

[amfern]

[0x000730d8] looks kinda like the voltage while the SoC VID is configured to 1.200v in bios
1.550 - (0.00625 * 0x3bc10000 >> 24) = 1.18125 (idle)
1.550 - (0.00625 * 0x34010000 >> 24) = 1.225 (load)

But if I set the SoC VID to 1.020v in bios, it doesn't really change by much.
1.550 - (0.00625 * 0x3c810000 >> 24) = 1.175 (idle)
1.550 - (0.00625 * 0x34410000 >> 24) = 1.225 (load)

[cyring]

What I can tell about my 3950X + X570 board is that Voltage SoC just fluctuates about 0.01V

If registers you found are indeed voltages then it could be other components: Fabric, I/O die ?
Can you post screenshots of the BIOS all voltages to search for a match ?

[amfern]

What I can tell about my 3950X + X570 board is that Voltage SoC just fluctuates about 0.01V

Thank you, i was wondering what it i should be looking for.
I noticed in my bios that SOC Voltage will fluctuate by about 0.01 too.

I can really pinpoint the voltage reported by [0x000730d8] register though.

[cyring]

Fyi, Registers you found in Zen4 don't seem to work on Zen2. These are results on my 3950X:

# ./zencli smu 0x00073010
[0x00073010] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111

# ./zencli smu 0x000730d8
[0x000730d8] READ(smu) = 0xffffffff (4294967295)
   60   56   52   48   44   40   36   32   28   24   20   16   12   08   04   00
 0000 0000 0000 0000 0000 0000 0000 0000 1111 1111 1111 1111 1111 1111 1111 1111

[cyring]

We can try other ports for the HSMP protocol:

Please change at lines:

CoreFreq/amd_reg.h

Line 164 in a28ea7b

#define SMU_AMD_INDEX_PORT_F17H 0x60

and replace with:

#define SMU_AMD_INDEX_PORT_F17H 	0xc4
#define SMU_AMD_DATA_PORT_F17H		0xc8

rebuild all and load driver

make clean all
insmod ./corefreqk.ko

next check into your kernel log (dmesg) if you encounter a HSMP timeout or error ?

[amfern]

I see timeout message

[ 3400.073129] CoreFreq: PCI_HSMP_Mailbox(1) Timeout
[ 3400.203246] CoreFreq(8:24:-1): Processor [ AF_61] Architecture [Zen4/Raphael] SMT [32/32]

[cyring]

In the article of igorslab.de you'll find the link to the "Renesas Electronics Corporation. 2020. ISL62776 Multiphase PWM Regulator for AMD CPUs Using SVI2"

Once PDF downloaded, you will go to Table 5. Serial VID Codes

This match the 0x0000b401 and could confirmed the found Register 0x00073010 as a SVI2

Concerning SVI3 on Zen4, I have not found much info on renesas.com

[amfern]

Interesting, it does fit the voltage.

[cyring]

I gave a quick look to the Renesas datasheets: SVI3 can be the answer; it requires an I2C dialog, I believe some userspace tools can help to query VID.

[cyring]

@amfern Hello,

I'm still not giving up with this subject.

Can you try the AMD HSMP driver ?
CoreFreq is apparently doing the same protocol but just in case I forgot something.

Based on latest Linux kernel, make sure CONFIG_AMD_HSMP is set. Mine is m for module.

zgrep CONFIG_AMD_HSMP /proc/config.gz 
CONFIG_AMD_HSMP=m

If unset, build your own kernel: in make menuconfig go to AMD HSMP Driver to enable that driver.

Next, load the module and post the result from the kernel log.

modprobe amd_hsmp

If you reach the point where driver is correctly inserted, we will submit commands in subsequent post.

Thank you
CyrIng

[amfern]

Sorry been busy with personal life and had some weird issue of random freezes, and disabling c-states seems to fix it.

Thank you for not giving up, I will try your suggestion abit later.

[amfern]

I get the error message

[ 1401.421165] amd_hsmp: HSMP is not supported on Fam:19 model:61
[ 1401.421168] amd_hsmp: Or Is HSMP disabled in BIOS ?

[cyring]

Does this mean that AMD has definitely fuse HSMP off, with the exception of EPYC processors ?

[amfern]

I don't see the option to enable HSMP in the bios either.

[cyring]

I've been answered that the functional HSMP mailbox protocol in 3950X was a design mistake. Oops.

[cyring]

Meanwhile in this EPYC/Genoa output, we have confirmation that HSMP has worked for Version and some PL fields:

Processor                                     [AMD EPYC 9274F 24-Core Processor]
|- Firmware                                                        [ 71.121.0-5]
...
|- Thermal Design Power                                      Package   < Enable>
   |- Power Limit                                                PL1   <  400 W>
   |- Time Window                                                TW1   <   0 ns>
   |- Power Limit                                                PL2   <  400 W>

[cyring]

But the boosted frequency ratio from SMU on Genoa is inconsistent.
See this issue
Bit fields and/or addresses of Zen4 seem to be now undocumented.