Provided by: rocm-smi_5.7.0-1ubuntu1_amd64 
NAME
rocm-smi - a tool to monitor AMD accelerators and GPUs
SYNOPSIS
rocm-smi [-h] [-d DEVICE [DEVICE ...]] [--alldevices] [--showhw] [-a] [-i] [-v] [-e [EVENT ...]]
[--showdriverversion] [--showtempgraph] [--showfwinfo [BLOCK ...]] [--showmclkrange]
[--showmemvendor] [--showsclkrange] [--showproductname] [--showserial] [--showuniqueid]
[--showvoltagerange] [--showbus] [--showpagesinfo] [--showpendingpages] [--showretiredpages]
[--showunreservablepages] [-f] [-P] [-t] [-u] [--showmemuse] [--showvoltage] [-b] [-c] [-g] [-l]
[-M] [-m] [-o] [-p] [-S] [-s] [--showmeminfo TYPE [TYPE ...]] [--showpids [VERBOSE]]
[--showpidgpus [SHOWPIDGPUS ...]] [--showreplaycount] [--showrasinfo [SHOWRASINFO ...]]
[--showvc] [--showxgmierr] [--showtopo] [--showtopoaccess] [--showtopoweight] [--showtopohops]
[--showtopotype] [--showtoponuma] [--showenergycounter] [--shownodesbw] [--showcomputepartition]
[--shownpsmode] [-r] [--resetfans] [--resetprofile] [--resetpoweroverdrive] [--resetxgmierr]
[--resetperfdeterminism] [--resetcomputepartition] [--resetnpsmode] [--setclock TYPE LEVEL]
[--setsclk LEVEL [LEVEL ...]] [--setmclk LEVEL [LEVEL ...]] [--setpcie LEVEL [LEVEL ...]]
[--setslevel SCLKLEVEL SCLK SVOLT] [--setmlevel MCLKLEVEL MCLK MVOLT] [--setvc POINT SCLK SVOLT]
[--setsrange SCLKMIN SCLKMAX] [--setmrange MCLKMIN MCLKMAX] [--setfan LEVEL] [--setperflevel
LEVEL] [--setoverdrive %] [--setmemoverdrive %] [--setpoweroverdrive WATTS] [--setprofile
SETPROFILE] [--setperfdeterminism SCLK] [--setcomputepartition
{CPX,SPX,DPX,TPX,QPX,cpx,spx,dpx,tpx,qpx}] [--setnpsmode
{NPS1,NPS2,NPS4,NPS8,nps1,nps2,nps4,nps8}] [--rasenable BLOCK ERRTYPE] [--rasdisable BLOCK
ERRTYPE] [--rasinject BLOCK] [--gpureset] [--load FILE | --save FILE] [--autorespond RESPONSE]
[--loglevel LEVEL] [--json] [--csv]
DESCRIPTION
Radeon Open Compute Platform (ROCm) - System Management Interface (SMI) - Command Line Interface (CLI).
rocm-smi is the python reference implementation of a CLI, from AMD, over its C system management library.
This tool acts as a command line interface for manipulating and monitoring the amdgpu kernel, and is
intended to replace and deprecate the existing rocm_smi.py CLI tool. It uses Ctypes to call the
rocm_smi_lib API. Recommended: At least one AMD GPU with ROCm driver installed Required: ROCm SMI library
installed (librocm_smi64).
OPTIONS
Main:
-h, --help
show this help message and exit
--gpureset
Reset specified GPU (One GPU must be specified). This flag will attempt to reset the GPU for a
specified device. This will invoke the GPU reset through the kernel debugfs file
amdgpu_gpu_recover. Note that GPU reset will not always work, depending on the manner in which
the GPU is hung.
--load FILE
Load Clock, Fan, Performance and Profile settings from FILE
--save FILE
Save Clock, Fan, Performance and Profile settings to FILE
-d DEVICE [DEVICE ...], --device DEVICE [DEVICE ...]
Execute command on specified device
Display Options:
--alldevices
--showhw
Show Hardware details
-a, --showallinfo
Show Temperature, Fan and Clock values
Topology:
-i, --showid
Show GPU ID
-v, --showvbios
Show VBIOS version
-e [EVENT ...], --showevents [EVENT ...]
Show event list
--showdriverversion
Show kernel driver version. This flag will print out the AMDGPU module version for amdgpu-pro or
ROCK kernels. For other kernels, it will simply print out the name of the kernel (uname).
--showtempgraph
Show Temperature Graph
--showfwinfo [BLOCK ...]
Show FW information
--showmclkrange
Show mclk range
--showmemvendor
Show GPU memory vendor
--showsclkrange
Show sclk range
--showproductname
Show SKU/Vendor name. This uses the pci.ids file to print out more information regarding the GPUs
on the system. update-pciids(8) may need to be executed on the machine to get the latest PCI ID
snapshot, as certain newer GPUs will not be present in the stock pci.ids file, and the file may
even be absent on certain OS installation types.
--showserial
Show GPU's Serial Number. This flag will print out the serial number for the graphics card.
NOTE: This is currently only supported on Vega20 server cards that support it. Consumer cards and
cards older than Vega20 will not support this feature.
--showuniqueid
Show GPU's Unique ID
--showvoltagerange
Show voltage range
--showbus
Show PCI bus number
Pages information:
--showpagesinfo
Show retired, pending and unreservable pages
--showpendingpages
Show pending retired pages
--showretiredpages
Show retired pages
--showunreservablepages
Show unreservable pages. The above four flags display the different "bad pages" as reported by
the kernel. The three types of pages are: Retired pages (reserved pages) - These pages are
reserved and are unable to be used. Pending pages - These pages are pending for reservation, and
will be reserved/retired. Unreservable pages - These pages are not reservable for some reason.
Hardware-related information:
-f, --showfan
Show current fan speed
-P, --showpower
Show current Average Graphics Package Power Consumption. "Graphics Package" refers to the GPU
plus any HBM (High-Bandwidth memory) modules, if present.
-t, --showtemp
Show current temperature
-u, --showuse
Show current GPU use
--showmemuse
Show current GPU memory used. This used to indicate how busy the respective blocks are. For
example, for --showuse (gpu_busy_percent sysfs file), the SMU samples every ms or so to see if any
GPU block (RLC, MEC, PFP, CP) is busy. If so, that's 1 (or high). If not, that's 0 (low). If we
have 5 high and 5 low samples, that means 50% utilization (50% GPU busy, or 50% GPU use). The
windows and sampling vary from generation to generation, but that is how GPU and VRAM use is
calculated in a generic sense. --showmeminfo (and VRAM% in concise output) will show the amount
of VRAM used (visible, total, GTT), as well as the total available for those partitions. The
percentage shown there indicates the amount of used memory in terms of current allocations.
--showvoltage
Show current GPU voltage.
Software-related/controlled information:
-b, --showbw
Show estimated PCIe use This shows an approximation of the number of bytes received and sent by
the GPU over the last second through the PCIe bus. Note that this will not work for APUs since
data for the GPU portion of the APU goes through the memory fabric and does not 'enter/exit' the
chip via the PCIe interface, thus no accesses are generated, and the performance counters can't
count accesses that are not generated. NOTE: It is not possible to easily grab the size of every
packet that is transmitted in real time, so the kernel estimates the bandwidth by taking the
maximum payload size (mps), which is the max size that a PCIe packet can be. and multiplies it by
the number of packets received and sent. This means that the SMI will report the maximum estimated
bandwidth, the actual usage could (and likely will be) less.
-c, --showclocks
Show current clock frequencies
┌────────────┬───────────────────────────────────────────────────────────────────────────────────┐
│ Clock type │ Description │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ DCEFCLK │ DCE (Display) │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ FCLK │ Data fabric (VG20 and later) - Data flow from XGMI, Memory, PCIe │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ SCLK │ GFXCLK (Graphics core) │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ Note │ SOCCLK split from SCLK as of Vega10. Pre-Vega10 they were both controlled by SCLK │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ MCLK │ GPU Memory (VRAM) │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ PCLK │ PCIe bus │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ Note │ This gives 2 speeds, PCIe Gen1 x1 and the highest available based on the hardware │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ SOCCLK │ System clock (VG10 and later) - DF, MM HUB, AT HUB, SYSTEM HUB, OSS, DFD │
├────────────┼───────────────────────────────────────────────────────────────────────────────────┤
│ Note │ DF split from SOCCLK as of Vega20. Pre-Vega20 they were both controlled by SOCCLK │
└────────────┴───────────────────────────────────────────────────────────────────────────────────┘
-g, --showgpuclocks
Show current GPU clock frequencies
-l, --showprofile
Show Compute Profile attributes
-M, --showmaxpower
Show maximum graphics package power this GPU will consume. This limit is enforced by the
hardware.
-m, --showmemoverdrive
Show current GPU Memory Clock OverDrive level
-o, --showoverdrive
Show current GPU Clock OverDrive level
-p, --showperflevel
Show current DPM Performance Level
-S, --showclkvolt
Show supported GPU and Memory Clocks and Voltages
-s, --showclkfrq
Show supported GPU and Memory Clock
--showmeminfo TYPE [TYPE ...]
Show Memory usage information for given block(s) TYPE This allows the user to see the amount of
used and total memory for a given block (vram, vis_vram, gtt). It returns the number of bytes
used and total number of bytes for each block 'all' can be passed as a field to return all blocks,
otherwise a quoted-string is used for multiple values (e.g. "vram vis_vram") vram refers to the
Video RAM, or graphics memory, on the specified device vis_vram refers to Visible VRAM, which is
the CPU-accessible video memory on the device gtt refers to the Graphics Translation Table.
--showpids [VERBOSE]
Show current running KFD PIDs (pass details to VERBOSE for detailed information)
--showpidgpus [SHOWPIDGPUS ...]
Show GPUs used by specified KFD PIDs (all if no arg given)
--showreplaycount
Show PCIe Replay Count
--showrasinfo [SHOWRASINFO ...]
Show RAS enablement information and error counts for the specified block(s) (all if no arg given)
This shows the RAS information for a given block. This includes enablement of the block
(currently GFX, SDMA and UMC are the only supported blocks) and the number of errors ue -
Uncorrectable errors ce - Correctable errors.
--showvc
Show voltage curve
--showxgmierr
Show XGMI error information since last read
--showtopo
Show hardware topology information
--showtopoaccess
Shows the link accessibility between GPUs
--showtopoweight
Shows the relative weight between GPUs
--showtopohops
Shows the number of hops between GPUs
--showtopotype
Shows the link type between GPUs
--showtoponuma
Shows the numa nodes
--showenergycounter
Energy accumulator that stores amount of energy consumed
--shownodesbw
Shows the numa nodes
--showcomputepartition
Shows current compute partitioning
--shownpsmode
Shows current NPS mode
Set options:
--setclock TYPE LEVEL
Set Clock Frequency Level(s) for specified clock (requires manual Perf level)
--setsclk LEVEL [LEVEL ...]
Set GPU Clock Frequency Level(s) (requires manual Perf level)
--resetperfdeterminism
Disable performance determinism
--setmclk LEVEL [LEVEL ...]
Set GPU Memory Clock Frequency Level(s) (requires manual Perf level)
The two above options allow you to set a mask for the levels. For example, if a GPU has 8 clock
levels, you can set a mask to use levels 0, 5, 6 and 7 with --setsclk 0 5 6 7 . This will only use
the base level, and the top 3 clock levels. This will allow you to keep the GPU at base level when
there is no GPU load, and the top 3 levels when the GPU load increases.
NOTES:
The clock levels will change dynamically based on GPU load based on the default
Compute and Graphics profiles. The thresholds and delays for a custom mask cannot
be controlled through the SMI tool.
This flag automatically sets the Performance Level to "manual" as the mask is not
applied when the Performance level is set to auto.
--setclock LEVEL LEVEL
Set Clock Frequency Level(s) for specified clock (requires manual Perf level)
--setpcie LEVEL [LEVEL ...]
Set PCIE Clock Frequency Level(s) (requires manual Perf level)
--setslevel SCLKLEVEL SCLK SVOLT
Change GPU Clock frequency (MHz) and Voltage (mV) for a specific Level
--setmlevel MCLKLEVEL MCLK MVOLT
Change GPU Memory clock frequency (MHz) and Voltage for (mV) a specific Level
--setvc POINT SCLK SVOLT
Change SCLK Voltage Curve (MHz mV) for a specific point
--setsrange SCLKMIN SCLKMAX
Set min and max SCLK speed
--setmrange MCLKMIN MCLKMAX
Set min and max MCLK speed
--setfan LEVEL
Set GPU Fan Speed (Level or %). This sets the fan speed to a value ranging from 0 to maxlevel, or
from 0%-100% If the level ends with a %, the fan speed is calculated as pct*maxlevel/100
(maxlevel is usually 255, but is determined by the ASIC).
NOTE: While the hardware is usually capable of overriding this value when required, it is
recommended to not set the fan level lower than the default value for extended periods
of time.
--setperflevel LEVEL
Set Performance Level This lets you use the pre-defined Performance Level values for clocks and
power profile, which can include: auto (Automatically change values based on GPU workload) low
(Keep values low, regardless of workload) high (Keep values high, regardless of workload) manual
(Only use values defined by --setsclk and --setmclk).
--setoverdrive %
Set GPU OverDrive level (requires manual|high Perf level)
--setmemoverdrive %
Set GPU Memory Overclock OverDrive level (requires manual|high Perf level) The above two options
are DEPRECATED IN NEWER KERNEL VERSIONS (use --setslevel/--setmlevel instead). This sets the
percentage above maximum for the max Performance Level. For example, --setoverdrive 20 will
increase the top sclk level by 20%, similarly --setmemoverdrive 20 will increase the top mclk
level by 20%. Thus if the maximum clock level is 1000MHz, then --setoverdrive 20 will increase
the maximum clock to 1200MHz. Note this option can be used in conjunction with the
--setsclk/--setmclk mask. Operating the GPU outside of specifications can cause irreparable
damage to your hardware. Please observe the warning displayed when using this option. This flag
automatically sets the clock to the highest level, as only the highest level is increased by the
OverDrive value.
--setpoweroverdrive WATTS
Set the maximum GPU power using Power OverDrive in Watts This allows users to change the maximum
power available to a GPU package. The input value is in Watts. This limit is enforced by the
hardware, and some cards allow users to set it to a higher value than the default that ships with
the GPU. This Power OverDrive mode allows the GPU to run at higher frequencies for longer periods
of time, though this may mean the GPU uses more power than it is allowed to use per power supply
specifications. Each GPU has a model-specific maximum Power OverDrive that is will take;
attempting to set a higher limit than that will cause this command to fail. Note operating the
GPU outside of specifications can cause irreparable damage to your hardware. Please observe the
warning displayed when using this option.
--setprofile SETPROFILE
Specify Power Profile level (#) or a quoted string of CUSTOM Profile attributes "# # # #..."
(requires manual Perf level) The Compute Profile accepts 1 or n parameters, either the Profile to
select (see --showprofile for a list of preset Power Profiles) or a quoted string of values for
the CUSTOM profile. Note these values can vary based on the ASIC, and may include:
SCLK_PROFILE_ENABLE - Whether or not to apply the 3 following SCLK settings (0=disable,1=enable).
Note: this is a hidden field. If set to 0, the following 3 values are displayed as '-’.
┌───────────────────┬────────────────────────────────────────────────────┐
│ Setting │ Description │
├───────────────────┼────────────────────────────────────────────────────┤
│ SCLK_UP_HYST │ Delay before sclk is increased (in milliseconds) │
├───────────────────┼────────────────────────────────────────────────────┤
│ SCLK_DOWN_HYST │ Delay before sclk is decresed (in milliseconds) │
├───────────────────┼────────────────────────────────────────────────────┤
│ SCLK_ACTIVE_LEVEL │ Workload required before sclk levels change (in %) │
└───────────────────┴────────────────────────────────────────────────────┘
MCLK_PROFILE_ENABLE - Whether or not to apply the 3 following MCLK settings (0=disable,1=enable)
NOTE: This is a hidden field. If set to 0, the following 3 values are displayed as '-'.
┌───────────────────┬────────────────────────────────────────────────────┐
│ Setting │ Description │
├───────────────────┼────────────────────────────────────────────────────┤
│ MCLK_UP_HYST │ Delay before mclk is increased (in milliseconds) │
├───────────────────┼────────────────────────────────────────────────────┤
│ MCLK_DOWN_HYST │ Delay before mclk is decresed (in milliseconds) │
├───────────────────┼────────────────────────────────────────────────────┤
│ MCLK_ACTIVE_LEVEL │ Workload required before mclk levels change (in %) │
└───────────────────┴────────────────────────────────────────────────────┘
Other settings:
┌──────────────────┬───────────────────────────────────────────────────────────────────────────┐
│ Setting │ Description │
├──────────────────┼───────────────────────────────────────────────────────────────────────────┤
│ BUSY_SET_POINT │ Threshold for raw activity level before levels change │
├──────────────────┼───────────────────────────────────────────────────────────────────────────┤
│ FPS │ Frames Per Second │
├──────────────────┼───────────────────────────────────────────────────────────────────────────┤
│ USE_RLC_BUSY │ When set to 1, DPM is switched up as long as RLC busy message is received │
├──────────────────┼───────────────────────────────────────────────────────────────────────────┤
│ MIN_ACTIVE_LEVEL │ Workload required before levels change (in %) │
└──────────────────┴───────────────────────────────────────────────────────────────────────────┘
NOTES:
When a compute queue is detected, the COMPUTE Power Profile values will be automatically
applied to the system, provided that the Perf Level is set to "auto".
The CUSTOM Power Profile is only applied when the Performance Level is set to "manual"
so using this flag will automatically set the performance level to "manual".
It is not possible to modify the non-CUSTOM Profiles. These are hard-coded by the kernel.
--setperfdeterminism SCLK
Set clock frequency limit to get minimal performance variation
--setcomputepartition {CPX,SPX,DPX,TPX,QPX,cpx,spx,dpx,tpx,qpx}
Set compute partition
--setnpsmode {NPS1,NPS2,NPS4,NPS8,nps1,nps2,nps4,nps8}
Set nps mode
--rasenable BLOCK ERRTYPE
Enable RAS for specified block and error type
--rasdisable BLOCK ERRTYPE
Disable RAS for specified block and error type
--rasinject BLOCK
Inject RAS poison for specified block (ONLY WORKS ON UNSECURE BOARDS)
Reset options:
-r, --resetclocks
Reset clocks and OverDrive to default
--resetfans
Reset fans to automatic (driver) control
--resetprofile
Reset Power Profile back to default
--resetpoweroverdrive
Set the maximum GPU power back to the device default state
--resetxgmierr
Reset XGMI error count
--resetperfdeterminism
Disable performance determinism
--resetcomputepartition
Resets to boot compute partition state
--resetnpsmode
Resets to boot NPS mode state
Auto-response options:
--autorespond RESPONSE
Response to automatically provide for all prompts (NOT RECOMMENDED)
Output options:
--loglevel LEVEL
This will allow the user to set a logging level for the SMI's actions, one of
debug/info/warning/error/critical. Currently this is only implemented for sysfs writes, but can
easily be expanded upon in the future to log other things from the SMI.
--json Print output in JSON format
--csv Print output in CSV format
OVERDRIVE SETTINGS
Enabling OverDrive requires both a card that support OverDrive and a driver parameter that enables its
use. Because OverDrive features can damage your card, most workstation and server GPUs cannot use
OverDrive. Consumer GPUs that can use OverDrive must enable this feature by setting bit 14 in the amdgpu
driver's ppfeaturemask module parameter.
For OverDrive functionality, the OverDrive bit (bit 14) must be enabled (by default, the OverDrive bit is
disabled on the ROCK and upstream kernels). This can be done by setting amdgpu.ppfeaturemask accordingly
in the kernel parameters, or by changing the default value inside amdgpu_drv.c (if building your own
kernel).
As an example, if the ppfeaturemask is set to 0xffffbfff (11111111111111111011111111111111), then
enabling the OverDrive bit would make it 0xffffffff (11111111111111111111111111111111).
These are the flags that require OverDrive functionality to be enabled for the flag to work:
--showclkvolt --showvoltagerange --showvc --showsclkrange --showmclkrange --setslevel --setmlevel
--setoverdrive --setpoweroverdrive --resetpoweroverdrive --setvc --setsrange --setmrange
DISCLAIMER
The information contained herein is for informational purposes only, and is subject to change without
notice. While every precaution has been taken in the preparation of this document, it may contain
technical inaccuracies, omissions and typographical errors, and AMD is under no obligation to update or
otherwise correct this information. Advanced Micro Devices, Inc. makes no representations or warranties
with respect to the accuracy or completeness of the contents of this document, and assumes no liability
of any kind, including the implied warranties of noninfringement, merchantability or fitness for
particular purposes, with respect to the operation or use of AMD hardware, software or other products
described herein.
COPYRIGHT
Copyright (c) 2014-2022 Advanced Micro Devices, Inc. All rights reserved.
The present manpage has been aggregated from the help output of rocm-smi and the readme github page, by
Maxime Chambonnet. This work is made available under the Expat license.
VERSION
1.4.1
The SMI will report a "version" which is the version of the kernel installed: uname. For ROCk
installations, this will be the AMDGPU module version (e.g. 5.0.71) For non-ROCk or monolithic ROCk
installations, this will be the kernel version, which will be equivalent to the following bash command:
uname -a | cut -d ' ' -f 3
BUGS
Please report bugs to rocm.smi.lib@amd.com, and in last resort to debian-ai@lists.debian.org .
AUTHORS
AMD Research and AMD HSA Software Development
Advanced Micro Devices, Inc.
www.amd.com
SEE ALSO
The full local documentation for the C rocm-smi library is available with the binary deb package librocm-
smi-dev, and is installed at: /usr/share/doc/librocm-smi-dev/ROCm_SMI_Manual.pdf .
The documentation for rocm-smi is maintained as a README markdown file at
https://github.com/RadeonOpenCompute/rocm_smi_lib/blob/master/python_smi_tools/README.md .
rocm-smi 1.4.1 2022-09-17 ROC-SMI(1)