embeded/jetson2026. 7. 27. 12:46

nvidia jetson agx thor / mig

mig 사용하기 위한 설정들 시작!

필요없는 데몬들 죽이고, persistent mode on

 

nvidia-smi -mig 1 은 root 권한이 있어야 한다.

아무튼 켜주고 나면 MIG M. 에 Disabled에서 Enabled로 변경된다.

 

grahic instance만 생성하고 compute instance는 안했는데

gi가 생성되어서 그런가  insufficient resource라고 에러가 난다.

nvidia@nvidia:~$ sudo nvidia-smi mig -cgi 80,83
Successfully created GPU instance ID  2 on GPU  0 using profile MIG 1g.0gb (ID 80)
Successfully created GPU instance ID  1 on GPU  0 using profile MIG 2g.0gb+gfx (ID 83)
nvidia@nvidia:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
nvidia@nvidia:~$ nvidia-smi
...
|  No MIG devices found                                                                   |
...
nvidia@nvidia:~$ sudo nvidia-smi mig -cci 80,83
Unable to create a compute instance on GPU  0 GPU instance ID  2 using profile 80: Invalid Argument
Failed to create compute instances: Invalid Argument

nvidia@nvidia:~$ sudo nvidia-smi mig -cgi 80,83 -C
Unable to create a GPU instance on GPU  0 using profile 80: Insufficient Resources
Failed to create GPU instances: Insufficient Resources

 

하드웨어 건드리는거라 그런가 은근히 root 권한 많이 요청하네

mig -dgi 옵션으로 삭제

nvidia@nvidia:~$ nvidia-smi mig -dgi
No GPU instances found: Insufficient Permissions
nvidia@nvidia:~$ sudo nvidia-smi mig -dgi
Successfully destroyed GPU instance ID  2 from GPU  0
Successfully destroyed GPU instance ID  1 from GPU  0

 

-C 주면 잘 생성된다.

nvidia@nvidia:~$ sudo nvidia-smi mig -cgi 80,83 -C
Successfully created GPU instance ID  2 on GPU  0 using profile MIG 1g.0gb (ID 80)
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  2 using profile MIG 1g.0gb (ID  0)
Successfully created GPU instance ID  1 on GPU  0 using profile MIG 2g.0gb+gfx (ID 83)
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  1 using profile MIG 2g.0gb (ID  1)
nvidia@nvidia:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
  MIG 2g.0gb      Device  0: (UUID: MIG-031bd8db-02a6-5420-968d-7fcdf907f7e4)
  MIG 1g.0gb      Device  1: (UUID: MIG-60591740-f26d-525b-ad5a-5f5ba115bc71)

 

옵션 주면 ci랑 di 생성된게 보인다.

 

-C 안주고 하나씩 하는거 어떻게 해야하지 ? 일단 맨땅에 헤딩 시작!

자원부족이라고 난리나고 있다.

nvidia@nvidia:~$ sudo nvidia-smi mig -dci
Successfully destroyed compute instance ID  0 from GPU  0 GPU instance ID  2
Successfully destroyed compute instance ID  0 from GPU  0 GPU instance ID  1
nvidia@nvidia:~$ sudo nvidia-smi mig -cgi 80,83 
Unable to create a GPU instance on GPU  0 using profile 80: Insufficient Resources
Failed to create GPU instances: Insufficient Resources
nvidia@nvidia:~$ sudo nvidia-smi mig -lci
No compute instances found: Not Found

 

먼가 빼먹은거 같으니 일단 삭제하고

nvidia@nvidia:~$ sudo nvidia-smi mig -dgi 80,83 
Option "80,83" is not recognized.
nvidia@nvidia:~$ sudo nvidia-smi mig -dgi 1,2
Option "1,2" is not recognized.
nvidia@nvidia:~$ sudo nvidia-smi mig -dgi
Successfully destroyed GPU instance ID  2 from GPU  0
Successfully destroyed GPU instance ID  1 from GPU  0

 

list ci/gi profile 명령으로 식별자 확인.. 음.. 위에서 not recognized 뜰만했네

그나저나 lcip 에서 1c.2g.0gb는 어따 써야 쓸 수 있을까?

 

몇 번 해보니 gi를 생성하고 거기에 ci 프로필을 연결해서 생성하면 되는것 같다.

nvidia@nvidia:~$ sudo nvidia-smi mig -cci -gi 1
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  1 using profile MIG 2g.0gb (ID  1)
nvidia@nvidia:~$ sudo nvidia-smi mig -cci -gi 2
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  2 using profile MIG 1g.0gb (ID  0)

 

-C 준 것 처럼 생성된 것 같기도?

 

그래서 동시에 접속해서 해보니 성능도 많이 떨어지지만, (기준 140 token/s <, 현재 60+ 45 token/s, 75%)

동시에 되는거 봐서는 잘 나눠진거 같은데..

문제는 소비전력이 이상하게 낮고(단독으로 돌리면 24W. 성능 저하랑 고라혀면 24*0.75 = 16W 딱인가?)

사용율은 N/A가 되서 나오지 않는 점. 그게 관리 상의 난점이군.

 

+

다시 한번 시도. 직접 지정해주는게 큰 의미는 없어 보이니까, 그냥 -C 하는게 나을듯 하다.

그나저나 ci 에서 0과 0*의 차이를 모르겠다. "*" 달린 애가 우선적으로 된다는 의미이려나?

nvidia@nvidia:~/llm$ sudo nvidia-smi mig -lcip
+--------------------------------------------------------------------------------------+
| Compute instance profiles:                                                           |
| GPU     GPU       Name             Profile  Instances   Exclusive       Shared       |
|       Instance                       ID     Free/Total     SM       DEC   ENC   OFA  |
|         ID                                                          CE    JPEG       |
|======================================================================================|
|   0      2       MIG 1g.0gb           0*     1/1            6        1     1     0   |
|                                                                      1     1         |
+--------------------------------------------------------------------------------------+
|   0      1       MIG 1c.2g.0gb        0      1/1            6        1     1     0   |
|                                                                      1     1         |
+--------------------------------------------------------------------------------------+
|   0      1       MIG 2g.0gb           1*     1/1           12        1     1     0   |
|                                                                      1     1         |
+--------------------------------------------------------------------------------------+

nvidia@nvidia:~/llm$ nvidia-smi -L; sudo nvidia-smi mig -lgi ; sudo nvidia-smi mig -lci
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
+---------------------------------------------------------+
| GPU instances:                                          |
| GPU   Name               Profile  Instance   Placement  |
|                            ID       ID       Start:Size |
|=========================================================|
|   0  MIG 1g.0gb            80        2          2:1     |
+---------------------------------------------------------+
|   0  MIG 2g.0gb+gfx        83        1          0:2     |
+---------------------------------------------------------+
No compute instances found: Not Found

nvidia@nvidia:~/llm$ sudo nvidia-smi mig -gi 2 -cci 1
Unable to create a compute instance on GPU  0 GPU instance ID  2 using profile 1: Invalid Argument
Failed to create compute instances: Invalid Argument

nvidia@nvidia:~/llm$ sudo nvidia-smi mig -gi 2 -cci 0
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  2 using profile MIG 1g.0gb (ID  0)

nvidia@nvidia:~/llm$ nvidia-smi -L; sudo nvidia-smi mig -lgi ; sudo nvidia-smi mig -lci 
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
  MIG 1g.0gb      Device  0: (UUID: MIG-60591740-f26d-525b-ad5a-5f5ba115bc71)
+---------------------------------------------------------+
| GPU instances:                                          |
| GPU   Name               Profile  Instance   Placement  |
|                            ID       ID       Start:Size |
|=========================================================|
|   0  MIG 1g.0gb            80        2          2:1     |
+---------------------------------------------------------+
|   0  MIG 2g.0gb+gfx        83        1          0:2     |
+---------------------------------------------------------+
+--------------------------------------------------------------------+
| Compute instances:                                                 |
| GPU     GPU       Name             Profile   Instance   Placement  |
|       Instance                       ID        ID       Start:Size |
|         ID                                                         |
|====================================================================|
|   0      2       MIG 1g.0gb           0         0          0:1     |
+--------------------------------------------------------------------+

 

저작자표시 (새창열림)

'embeded > jetson' 카테고리의 다른 글

nvidia jetson agx thor / 디버그 usb  (0) 2026.07.27
nvidia-smi mig  (0) 2026.07.26
nvidia jetson agx thor dev kit 설치 - USB boot  (0) 2026.07.24
nvidia MIG(multiple instance GPU) - thor  (0) 2026.07.24
nvidia jetson 아니고 nvidia thor  (0) 2026.07.24
Posted by 구차니
embeded/jetson2026. 7. 27. 12:36

nvidia jetson agx thor / 디버그 usb

한번 꽂아보니 4개 포트가 뜨고

[250997.559440] usb 1-2: new full-speed USB device number 11 using xhci_hcd
[250997.687952] usb 1-2: New USB device found, idVendor=0955, idProduct=7045, bcdDevice= 0.01
[250997.687971] usb 1-2: New USB device strings: Mfr=1, Product=2, SerialNumber=3
[250997.687978] usb 1-2: Product: Tegra On-Platform Operator
[250997.687983] usb 1-2: Manufacturer: NVIDIA
[250997.687988] usb 1-2: SerialNumber: TOPOA735A12B
[250997.698054] hid-generic 0003:0955:7045.000B: hiddev2,hidraw9: USB HID v1.10 Device [NVIDIA Tegra On-Platform Operator] on usb-0000:00:14.0-2/input0
[250997.736401] cdc_acm 1-2:1.1: ttyACM0: USB ACM device
[250997.736738] cdc_acm 1-2:1.3: ttyACM1: USB ACM device
[250997.737020] cdc_acm 1-2:1.5: ttyACM2: USB ACM device
[250997.737348] cdc_acm 1-2:1.7: ttyACM3: USB ACM device
[250997.737379] usbcore: registered new interface driver cdc_acm
[250997.737381] cdc_acm: USB Abstract Control Model driver for USB modems and ISDN adapters

 

/dev/ttyACM0 (왼쪽) 은 커널 로그

/dev/ttyACM1 (오른쪽)은 uefi 가 뜬다.

 

esc 누르니까 저렇게 뜬다. 예전에 imx8mp evk 보드에서 windows on arm 돌리는 느낌이네..

선택이 안보여서 minicom -c on 옵션주고 나니 이제야 머가 선택되었는지 보인다.

저작자표시 (새창열림)

'embeded > jetson' 카테고리의 다른 글

nvidia jetson agx thor / mig  (0) 2026.07.27
nvidia-smi mig  (0) 2026.07.26
nvidia jetson agx thor dev kit 설치 - USB boot  (0) 2026.07.24
nvidia MIG(multiple instance GPU) - thor  (0) 2026.07.24
nvidia jetson 아니고 nvidia thor  (0) 2026.07.24
Posted by 구차니
embeded/jetson2026. 7. 26. 22:13

nvidia-smi mig

아무생각없이 블로그 글 따라보고

-cgi -C 했는데

원래 대로라면

-cgi

-cci 할걸

-cgi -C로 해결하는 듯?

 

$ nvidia-smi mig

    mig -- Multi Instance GPU management.

    Usage: nvidia-smi mig [options]

    Options include:
    [-h | --help]: Display help information.
    [-i | --id]: Enumeration index, PCI bus ID or UUID.
                 Provide comma separated values for more than one device.
    [-gi | --gpu-instance-id]: GPU instance ID.
                               Provide comma separated values for more than one GPU instance.
    [-ci | --compute-instance-id]: Compute instance ID.
                                   Provide comma separated values for more than one compute
                                   instance.
    [-lgip | --list-gpu-instance-profiles]: List supported GPU instance profiles.
                                            Option -i can be used to restrict the command to
                                            run on a specific GPU.
    [-lgipp | --list-gpu-instance-possible-placements]: List possible GPU instance placements
                                                        in the following format, {Start}:Size.
                                                        Option -i can be used to restrict the
                                                        command to run on a specific GPU.
    [-C | --default-compute-instance]: Create compute instance with the default profile when used
                                       with the option to create a GPU instance (-cgi).
    [-cgi | --create-gpu-instance]: Create GPU instances for the given profile tuples. A profile
                                    tuple consists of a profile name or ID and an optional placement
                                    specifier, which consists of a colon and a placement start index.
                                    Provide comma separated values for more than one profile tuple.
                                    Option -i can be used to restrict the command to run on
                                    a specific GPU.
    [-dgi | --destroy-gpu-instance]: Destroy GPU instances.
                                     Options -i and -gi can be used individually or combined
                                     to restrict the command to run on a specific GPU or GPU
                                     instance.
    [-lgi | --list-gpu-instances]: List GPU instances.
                                   Option -i can be used to restrict the command to run on a
                                   specific GPU.
    [-lcip | --list-compute-instance-profiles]: List supported compute instance profiles.
                                                Options -i and -gi can be used individually or
                                                combined to restrict the command to run on a
                                                specific GPU or GPU instance.
    [-lcipp | --list-compute-instance-possible-placements]: List possible compute instance placements
                                                            in the following format, {Start}:Size.
                                                            Options -i and -gi can be used individually or
                                                            combined to restrict the command to run on a
                                                            specific GPU or GPU instance.
    [-cci | --create-compute-instance]: Create compute instance for the given profile name or IDs.
                                        Provide comma separated values for more than one profile.
                                        If no profile name or ID is given, then the default*
                                        compute instance profile ID will be used. Options -i and
                                        -gi can be used individually or combined to restrict the
                                        command to run on a specific GPU or GPU instance.
    [-dci | --destroy-compute-instance]: Destroy compute instances.
                                         Options -i, -gi and -ci can be used individually or
                                         combined to restrict the command to run on a specific
                                         GPU or GPU instance or compute instance.
    [-lci | --list-compute-instances]: List compute instances.
                                       Options -i and -gi can be used individually or combined
                                       to restrict the command to run on a specific GPU or GPU
                                       instance.

저작자표시 (새창열림)

'embeded > jetson' 카테고리의 다른 글

nvidia jetson agx thor / mig  (0) 2026.07.27
nvidia jetson agx thor / 디버그 usb  (0) 2026.07.27
nvidia jetson agx thor dev kit 설치 - USB boot  (0) 2026.07.24
nvidia MIG(multiple instance GPU) - thor  (0) 2026.07.24
nvidia jetson 아니고 nvidia thor  (0) 2026.07.24
Posted by 구차니
개소리 왈왈/컴퓨터2026. 7. 26. 22:04

음.. 쓰레기를 샀나?

왼손 키보드 샀는데 망

블루투스로 핸드폰에서 연결해도 입력도 안되고

usb type a는 마우스 연결용 인 것 같고

usb type c는 전원 공급용 / 설정용 같은데 설정 프로그램이 없으면 먼가 안되는 것 같은데 

프로그램을 찾을수 없고.. 으아아

 

[링크 : https://www.lazada.com.ph/products/pinklehub-rk-b30-wireless-rainbow-keyboard-red-switch-rk-b30-red-switch-i2390144881.html]

저작자표시 (새창열림)

'개소리 왈왈 > 컴퓨터' 카테고리의 다른 글

usb-c 가 있다고 dp alt가 되는건 아니다  (0) 2026.07.14
반도체 수급불량이 sd 카드에도..  (0) 2026.07.08
socamm lpddr  (0) 2026.07.03
runpod 조사중  (0) 2026.07.01
cpu 줍줍  (0) 2026.06.30
Posted by 구차니
개소리 왈왈/직딩의 비애2026. 7. 25. 22:50

위 내시경

나라에서 내려주는(!) 2년에 한번 고문

아우.. 이거 안할수 있는 방법 없나!!!!!

저작자표시 (새창열림)

'개소리 왈왈 > 직딩의 비애' 카테고리의 다른 글

기사 / ai 성과 측정한다는 삼성전자  (0) 2026.06.17
PV5 첫 승차  (6) 2026.06.15
세상이 왜 이렇게 되었을까  (0) 2026.05.11
열풍기 잼나네  (6) 2026.04.20
쏘쏘  (0) 2026.04.15
Posted by 구차니
embeded/jetson2026. 7. 24. 15:54

nvidia jetson agx thor dev kit 설치 - USB boot

2026.07.24 기준

jetpack 7.2 / jetson linux 39.2

[링크 : https://developer.nvidia.com/embedded/jetpack/downloads]

[링크 : https://docs.nvidia.com/jetson/agx-thor-devkit/user-guide/latest/quick_start.html]

[링크 : https://docs.nvidia.com/jetson/agx-thor-devkit/user-guide/latest/twa_display_handoff.html]

 

usb 키보드 연결하고 esc 연타.

 

그나저나 먼가 익숙한(?) imx8mp 때의 부트로더라고 해야하나

UEFI 로더라고 해야하나 이게 보이네?!

boot manager 선택하고

 

NVIDIA Configuration 선택

 

Boot Configuration 선택

 

밑으로 내려가서(위로 한번 누르면 바로 내려감)

SOC Display Hand-Off Mode를 Auto 로 해주고 usb에 iso 구워서 켜면 그냥 pc 처럼 설치가 되나보다.

 

그냥 USB 꽂고 F11 눌러서 부팅 장치 선택하니(완전 PC네!)

USB 선택하고 엔터!

 

최후 통첩(?) 30초 기다려준다!

 

우분투 grub 는 깜박잊고(맨날 보던거라..) 그냥 넘어가버렸고

하위 선택에서 'Install on NVMe' 선택

 

많이 깔꺼니까 15분 정도 기다려! 라니!~ ㅋㅋ

 

엣! 잠시 자리 비운 사이에 리부팅 되었나 보다. 먼가 업데이트 하는 중

 

휴.. 조마조마했는데 다행이 먼가 뜬다!

 

먼가 설정하고 하니 24.04.4 기반이군

저작자표시 (새창열림)

'embeded > jetson' 카테고리의 다른 글

nvidia jetson agx thor / 디버그 usb  (0) 2026.07.27
nvidia-smi mig  (0) 2026.07.26
nvidia MIG(multiple instance GPU) - thor  (0) 2026.07.24
nvidia jetson 아니고 nvidia thor  (0) 2026.07.24
jetson nano + llama.cpp vulkan  (0) 2026.07.21
Posted by 구차니
개소리 왈왈/블로그2026. 7. 24. 14:59

불법촬영물 여부 검토

티스토리는 이거 대응한다고 업로드 하면 약간의 시간이 지난 후에 뜬다.

살색이 많은 이미지도 아닌데 왜 이래?

저작자표시 (새창열림)

'개소리 왈왈 > 블로그' 카테고리의 다른 글

다음은 언제이려나? ㅋㅋ  (2) 2026.06.23
한 달에 한 번?  (0) 2026.06.17
개인정보 유출 통지 - 데이원 컴퍼니 <- 콜로소  (2) 2026.06.11
잠깐의 행복했던 방문자 수  (0) 2026.06.08
도메인 결제  (4) 2026.06.01
Posted by 구차니
모종의 음모/CAN2026. 7. 24. 14:34

MCP2515 on rpi .. 2번째

벌써 3년전에 가볍게(?) 해봤었네..

또 까먹다니 이 기억력 어쩔.. ㅠㅠ

그때와 다른건.. 작년 부터 /boot/config.txt의 경로가 달라졌다는 정도?

$ uname -a
Linux raspberrypi 6.12.25+rpt-rpi-v7 #1 SMP Raspbian 1:6.12.25-1+rpt1 (2025-04-30) armv7l GNU/Linux

$ cat /boot/config.txt
DO NOT EDIT THIS FILE

The file you are looking for has moved to /boot/firmware/config.txt

$ cat /boot/firmware/config.txt
[all]
dtoverlay=mcp2515-can0,oscillator=8000000,interrupt=12
dtoverlay=spi-bcm2835-overlay

$ sudo ip link set can0 up type can bitrate 125000
$ ifconfig can0
can0: flags=193<UP,RUNNING,NOARP>  mtu 16
        unspec 00-00-00-00-00-00-00-00-00-00-00-00-00-00-00-00  txqueuelen 10  (UNSPEC)
        RX packets 0  bytes 0 (0.0 B)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 0  bytes 0 (0.0 B)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

$ dmesg | grep -i can
[   12.948060] CAN device driver interface
[   13.070005] mcp251x spi0.0 can0: MCP2515 successfully initialized.

$ lsmod |grep can
can_dev                40960  1 mcp251x

 

loopback 설정. ifconfig 상으로는 별다른 티가 안난다

$ sudo ip link set can0 up type can bitrate 125000 loopback on
$ sudo ifconfig can0
can0: flags=193<UP,RUNNING,NOARP>  mtu 16
        unspec 00-00-00-00-00-00-00-00-00-00-00-00-00-00-00-00  txqueuelen 10  (UNSPEC)
        RX packets 0  bytes 0 (0.0 B)
        RX errors 0  dropped 0  overruns 0  frame 0
        TX packets 0  bytes 0 (0.0 B)
        TX errors 0  dropped 0 overruns 0  carrier 0  collisions 0

[링크 : https://forums.raspberrypi.com/viewtopic.php?t=7027&start=225]

 

아무튼 loopback을 활성화 하고 ssh 로 창 두개 열어서 한쪽은 candump 실행해두고

다른 한쪽에서 cansend로 보내면 보내진다.

$ cansend can0 5A1#11.2233.44556677.88
$ candump can0
  can0  5A1   [8]  11 22 33 44 55 66 77 88
  can0  5A1   [8]  11 22 33 44 55 66 77 88

 

$ demsg | tail
[  965.109222] can: controller area network core
[  965.109370] NET: Registered PF_CAN protocol family
[  965.136117] can: raw protocol
[ 1028.671698] mcp251x spi0.0 can0: bus-off

[링크 : https://whiteknight3672.tistory.com/346]

[링크 : https://blog.naver.com/mts90/221644429413]

[링크 : https://wp.mytv.kr/2019/10/31/raspberry-pi-에서-can-통신-사용하기/] 결선도

2023.10.31 - [모종의 음모/CAN] - MCP2515 on rpi

2025.02.19 - [모종의 음모/CAN] - candump, cansend와 vcan

 

+

MCP2515 / TJA1050

 

D0695323C  라고 써있어서

이게 8 MHz 인지 16 MHz 인질 모르겠다. 검색해도 안나온다.

일단 두대를 보니 서로 다른게 써있어서 그냥 시리얼 번호라고 생각하는 중.

 

일단은 3.3V했는데 이러면 외부로 신호는 발생시키지 못할 듯.

 

예전에도 본건 데 전원이 잘못되었다고 개조를 해야 한다고 하는데

지금 다시 읽어보니 라즈베리 파이 gpio 가 3.3V 라서 아두이노와는 다르게 5V로 주면 라즈베리가 데미지 입게된다.

Because MCP2515 must be powered from 3,3V, because MCP2515 is directly connected to Pi GPIO and GPIO only accept 3,3V

[링크 : https://forums.raspberrypi.com/viewtopic.php?t=141052]

 

데이터 시트를 찾아보면

MCP2515는 2.7 V~5.5 V 면 되지만 SPI 도 Vdd 전압 따라가게 되면 흐음..

[링크 : https://ww1.microchip.com/downloads/en/DeviceDoc/MCP2515-Stand-Alone-CAN-Controller-with-SPI-20001801J.pdf]

 

CAN tranceiver칩은 4.75~5.25V 사이에 전압을 공급해야 한다.

[링크 : https://www.nxp.com/docs/en/data-sheet/TJA1050.pdf]

 

걍 속편하게 SPI 쪽만 레벨 시프터 쓸까? 개조가 나을까..

 

저작자표시 (새창열림)

'모종의 음모 > CAN' 카테고리의 다른 글

obd can  (0) 2026.07.23
DeviceNet CAN  (0) 2025.04.01
candump, cansend와 vcan  (0) 2025.02.19
linux can 테스트(가상 CAN IF)  (0) 2025.02.18
can-utils(cansend, candump)  (0) 2025.02.18
Posted by 구차니
프로그램 사용/ai 프로그램2026. 7. 24. 11:52

nvidia thor - LLM

모델이 작은거라 그런가 94% 인데 33W라.. 좀 더 빡세게 굴려봐야하나 ㅋㅋ

 

$ ./llama-b10099/llama-cli -m ./model/gemma4-e4b-qat/gemma-4-E4B-it-qat-UD-Q4_K_XL.gguf 

> 안녕?
[ Prompt: 8.0 t/s | Generation: 45.2 t/s ]

> 너에 대한 소개
[ Prompt: 92.5 t/s | Generation: 51.4 t/s ]

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
[ Prompt: 803.0 t/s | Generation: 48.9 t/s ]

 

$ ./llama-b10099/llama-cli -m ./model/qwen3.5/Qwen3.5-0.8B-UD-Q4_K_XL.gguf -rea off

> 안녕?  
[ Prompt: 152.3 t/s | Generation: 121.5 t/s ]

> 너에 대한 소개
[ Prompt: 218.0 t/s | Generation: 140.4 t/s ]

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
[ Prompt: 68.2 t/s | Generation: 136.5 t/s ]

 

컨텍스트 줄인다고 빨라지진 않네

$ ./llama-b10099/llama-cli -m ./model/qwen3.5/Qwen3.5-0.8B-UD-Q4_K_XL.gguf -rea off -c 4096

> 안녕?
[ Prompt: 177.3 t/s | Generation: 131.4 t/s ]

> 너에 대한 소개
[ Prompt: 335.9 t/s | Generation: 140.5 t/s ]

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
[ Prompt: 1389.9 t/s | Generation: 136.3 t/s ]

 

$ ./llama-b10099/llama-cli -m ./model/gemma4-e4b/gemma-4-E4B-it-Q4_K_M.gguf 
> 안녕?
[ Prompt: 1.0 t/s | Generation: 26.9 t/s ]

> 안녕?
[ Prompt: 65.8 t/s | Generation: 42.2 t/s ]

> 너에 대한 소개
[ Prompt: 121.2 t/s | Generation: 41.9 t/s ]

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
[ Prompt: 916.7 t/s | Generation: 41.1 t/s ]

 

gemma4-31B 는 조금 더 뜨끈해진다.

 

$ ./llama-b10099/llama-cli -m ./model/gemma4-31b/gemma-4-31B-it-UD-Q2_K_XL.gguf 

> 안녕?
[ Prompt: 1.1 t/s | Generation: 6.9 t/s ]

> 너에 대한 소개
[ Prompt: 11.5 t/s | Generation: 7.1 t/s ]

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
[ Prompt: 68.3 t/s | Generation: 6.9 t/s ]

 

메모리가 의외로 얼마 안먹나?

nvidia@localhost:~/llm$ free -h
               total        used        free      shared  buff/cache   available
Mem:           122Gi        40Gi       2.2Gi        26Mi        81Gi        82Gi
Swap:          2.0Gi       256Ki       2.0Gi
nvidia@localhost:~/llm$ free -h
               total        used        free      shared  buff/cache   available
Mem:           122Gi        37Gi       5.0Gi        26Mi        81Gi        85Gi
Swap:          2.0Gi       256Ki       2.0Gi

 

 파일이 11G 인데 저거밖에 안 먹는게 맞..나?

$ ll -h ./model/gemma4-31b/gemma-4-31B-it-UD-Q2_K_XL.gguf 
-rw-rw-r-- 1 nvidia nvidia 11G  7월 24 11:11 ./model/gemma4-31b/gemma-4-31B-it-UD-Q2_K_XL.gguf

 

qwen3.6 35B는 의외로 빠르다 머지?

$ ./llama-b10099/llama-cli -m ./model/qwen3.6_35B/Qwen3.6-35B-A3B-UD-Q2_K_XL.gguf 

> 안녕?
[ Prompt: 6.0 t/s | Generation: 43.9 t/s ]

> 너에 대한 소개
[ Prompt: 39.6 t/s | Generation: 45.1 t/s ]

> 파이썬으로 셀레니움을 통해 웹을 서칭하고 텍스트만 추출하고 makrdown 으로 변환후  md 파일과 pdf로 저장하는 기능을 구현해줘
[ Prompt: 93.7 t/s | Generation: 43.6 t/s ]

 

메모리는 봐도 모르겠다

nvidia@localhost:~/llm$ free -h
               total        used        free      shared  buff/cache   available
Mem:           122Gi        38Gi       3.8Gi        26Mi        81Gi        83Gi
Swap:          2.0Gi       256Ki       2.0Gi

nvidia@localhost:~/llm$ free -h
               total        used        free      shared  buff/cache   available
Mem:           122Gi        37Gi       5.0Gi        26Mi        81Gi        85Gi
Swap:          2.0Gi       256Ki       2.0Gi

 

+

$ ./llama-b10099/llama-cli -m ./model/gemma4-e4b/gemma-4-E4B-it-Q4_K_M.gguf

 

$ ./llama-b10099/llama-cli -m ./model/qwen3.6_35B/Qwen3.6-35B-A3B-UD-Q2_K_XL.gguf 

저작자표시 (새창열림)

'프로그램 사용 > ai 프로그램' 카테고리의 다른 글

llama.cpp for arm64  (0) 2026.07.20
qwen3.5 0.8B  (0) 2026.07.19
LLM 지식 증류(knowledge distilation), lora  (0) 2026.07.18
sLLM (경량 언어모델)  (0) 2026.07.18
딥러닝은 참.. 전기가 많이 드는구나..  (0) 2026.07.16
Posted by 구차니
embeded/jetson2026. 7. 24. 11:39

nvidia MIG(multiple instance GPU) - thor

귀찮아서 명령어 조합 저장

sudo systemctl stop gdm
sudo systemctl stop nvargus-daemon
sudo nvidia-smi -pm 1
sudo nvidia-smi -i 0 -mig 1
sudo nvidia-smi mig -cgi 80,83 -C

 

----

헛다리를 짚은것 같은데 메모리가 안나오는건 공유 메모리를 써서 인가.. 싶기도 하고?

아무튼 thor iGPU에서는 최대 2개로 나누어 쓸 수 있나 본데

gfx는 외부 출력 포트를 쓸수 있냐 없냐의 차이인것 같고

[링크 : https://docs.nvidia.com/datacenter/tesla/mig-user-guide/supported-mig-profiles.html]

 

nvidia@localhost:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)

nvidia@localhost:~$ nvidia-smi -i 0 -q

==============NVSMI LOG==============

Timestamp                                 : Fri Jul 24 11:33:07 2026
Driver Version                            : 580.00
CUDA Version                              : 13.0

Attached GPUs                             : 1
GPU 00000000:01:00.0
    Product Name                          : NVIDIA Thor
    Product Brand                         : GeForce
    Product Architecture                  : Blackwell
    Display Mode                          : Requested functionality has been deprecated
    Display Attached                      : No
    Display Active                        : Disabled
    Persistence Mode                      : Disabled
    Addressing Mode                       : ATS
    MIG Mode
        Current                           : Disabled
        Pending                           : Disabled
    Accounting Mode                       : Disabled
    Accounting Mode Buffer Size           : 4000
    Driver Model
        Current                           : N/A
        Pending                           : N/A
    Serial Number                         : N/A
    GPU UUID                              : GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600
    GPU PDI                               : 0x9257b02c0bd38a20
    Minor Number                          : 0
    VBIOS Version                         : N/A
    MultiGPU Board                        : No
    Board ID                              : 0x100
    Board Part Number                     : N/A
    GPU Part Number                       : 2B00--A1
    FRU Part Number                       : N/A
    Platform Info
        Chassis Serial Number             : N/A
        Slot Number                       : N/A
        Tray Index                        : N/A
        Host ID                           : N/A
        Peer Type                         : N/A
        Module Id                         : N/A
        GPU Fabric GUID                   : N/A
    Inforom Version
        Image Version                     : N/A
        OEM Object                        : N/A
        ECC Object                        : N/A
        Power Management Object           : N/A
    Inforom BBX Object Flush
        Latest Timestamp                  : N/A
        Latest Duration                   : N/A
    GPU Operation Mode
        Current                           : N/A
        Pending                           : N/A
    GPU C2C Mode                          : Disabled
    GPU Virtualization Mode
        Virtualization Mode               : None
        Host VGPU Mode                    : N/A
        vGPU Heterogeneous Mode           : N/A
    GPU Recovery Action                   : None
    GSP Firmware Version                  : 580.00
    IBMNPU
        Relaxed Ordering Mode             : N/A
    PCI
        Bus                               : 0x01
        Device                            : 0x00
        Domain                            : 0x0000
        Base Classcode                    : 0x3
        Sub Classcode                     : 0x2
        Device Id                         : 0x2B0010DE
        Bus Id                            : 00000000:01:00.0
        Sub System Id                     : 0x00000000
        GPU Link Info
            PCIe Generation
                Max                       : 1
                Current                   : 1
                Device Current            : 1
                Device Max                : 5
                Host Max                  : 1
            Link Width
                Max                       : 16x
                Current                   : 1x
        Bridge Chip
            Type                          : N/A
            Firmware                      : N/A
        Replays Since Reset               : 0
        Replay Number Rollovers           : 0
        Tx Throughput                     : N/A
        Rx Throughput                     : N/A
        Atomic Caps Outbound              : FETCHADD_32 FETCHADD_64 SWAP_32 SWAP_64 CAS_32 CAS_64 
        Atomic Caps Inbound               : N/A
    Fan Speed                             : N/A
    Performance State                     : N/A
    Clocks Event Reasons                  : N/A
    Clocks Event Reasons Counters
        SW Power Capping                  : N/A
        Sync Boost                        : N/A
        SW Thermal Slowdown               : N/A
        HW Thermal Slowdown               : N/A
        HW Power Braking                  : N/A
    Sparse Operation Mode                 : N/A
    FB Memory Usage
        Total                             : N/A
        Reserved                          : N/A
        Used                              : N/A
        Free                              : N/A
    BAR1 Memory Usage
        Total                             : N/A
        Used                              : N/A
        Free                              : N/A
    Conf Compute Protected Memory Usage
        Total                             : 0 MiB
        Used                              : 0 MiB
        Free                              : 0 MiB
    Compute Mode                          : Default
    Utilization
        GPU                               : 98 %
        Memory                            : 0 %
        Encoder                           : 0 %
        Decoder                           : 0 %
        JPEG                              : 0 %
        OFA                               : 0 %
    Encoder Stats
        Active Sessions                   : 0
        Average FPS                       : 0
        Average Latency                   : 0
    FBC Stats
        Active Sessions                   : 0
        Average FPS                       : 0
        Average Latency                   : 0
    DRAM Encryption Mode
        Current                           : N/A
        Pending                           : N/A
    ECC Mode
        Current                           : N/A
        Pending                           : N/A
    ECC Errors
        Volatile
            SRAM Correctable              : N/A
            SRAM Uncorrectable Parity     : N/A
            SRAM Uncorrectable SEC-DED    : N/A
            DRAM Correctable              : N/A
            DRAM Uncorrectable            : N/A
        Aggregate
            SRAM Correctable              : N/A
            SRAM Uncorrectable Parity     : N/A
            SRAM Uncorrectable SEC-DED    : N/A
            DRAM Correctable              : N/A
            DRAM Uncorrectable            : N/A
            SRAM Threshold Exceeded       : N/A
        Aggregate Uncorrectable SRAM Sources
            SRAM L2                       : N/A
            SRAM SM                       : N/A
            SRAM Microcontroller          : N/A
            SRAM PCIE                     : N/A
            SRAM Other                    : N/A
        Channel Repair Pending            : N/A
        TPC Repair Pending                : N/A
    Retired Pages
        Single Bit ECC                    : N/A
        Double Bit ECC                    : N/A
        Pending Page Blacklist            : N/A
    Remapped Rows                         : N/A
    Temperature
        GPU Current Temp                  : 60 C
        GPU T.Limit Temp                  : N/A
        GPU Shutdown Temp                 : N/A
        GPU Slowdown Temp                 : N/A
        GPU Max Operating Temp            : N/A
        GPU Target Temperature            : N/A
        Memory Current Temp               : N/A
        Memory Max Operating Temp         : N/A
    GPU Power Readings
        Average Power Draw                : N/A
        Instantaneous Power Draw          : 36.21 W
        Current Power Limit               : N/A
        Requested Power Limit             : N/A
        Default Power Limit               : N/A
        Min Power Limit                   : N/A
        Max Power Limit                   : N/A
    GPU Memory Power Readings 
        Average Power Draw                : N/A
        Instantaneous Power Draw          : N/A
    Module Power Readings
        Average Power Draw                : N/A
        Instantaneous Power Draw          : N/A
        Current Power Limit               : N/A
        Requested Power Limit             : N/A
        Default Power Limit               : N/A
        Min Power Limit                   : N/A
        Max Power Limit                   : N/A
    Power Smoothing                       : N/A
    Workload Power Profiles
        Requested Profiles                : N/A
        Enforced Profiles                 : N/A
    Clocks
        Graphics                          : 1575 MHz
        SM                                : N/A
        Memory                            : N/A
        Video                             : 315 MHz
    Applications Clocks
        Graphics                          : Requested functionality has been deprecated
        Memory                            : Requested functionality has been deprecated
    Default Applications Clocks
        Graphics                          : Requested functionality has been deprecated
        Memory                            : Requested functionality has been deprecated
    Deferred Clocks
        Memory                            : N/A
    Max Clocks
        Graphics                          : N/A
        SM                                : N/A
        Memory                            : N/A
        Video                             : N/A
    Max Customer Boost Clocks
        Graphics                          : N/A
    Clock Policy
        Auto Boost                        : N/A
        Auto Boost Default                : N/A
    Fabric
        State                             : Deprecated; See field Probe State instead
        Probe State                       : N/A
        Status                            : Deprecated; See field Probe Status instead
        Probe Status                      : N/A
        CliqueId                          : N/A
        ClusterUUID                       : N/A
        Health
            Summary                       : N/A
            Bandwidth                     : N/A
            Route Recovery in progress    : N/A
            Route Unhealthy               : N/A
            Access Timeout Recovery       : N/A
            Incorrect Configuration       : N/A
    Processes
        GPU instance ID                   : N/A
        Compute instance ID               : N/A
        Process ID                        : 3559
            Type                          : G
            Name                          : /usr/lib/xorg/Xorg
            Used GPU Memory               : 95 MiB
        GPU instance ID                   : N/A
        Compute instance ID               : N/A
        Process ID                        : 3775
            Type                          : G
            Name                          : /usr/bin/gnome-shell
            Used GPU Memory               : 58 MiB
        GPU instance ID                   : N/A
        Compute instance ID               : N/A
        Process ID                        : 8710
            Type                          : C+G
            Name                          : ./llama-b10099/llama-cli
            Used GPU Memory               : 33671 MiB
    Capabilities
        EGM                               : disabled

nvidia@localhost:~$ nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|

[링크 : https://blog.naver.com/techtrip/224167049409]

[링크 : https://docs.nvidia.com/jetson/archives/r39.2/DeveloperGuide/SD/MiG.html#getting-started-with-mig-on-jetson-thor]

+

gdm만 죽이나 isolate 해서 죽이나 별반 차이 없는 것 같이 안된다 ㅠㅠ

nvidia@localhost:~$ nvidia-smi -i 0
Fri Jul 24 11:57:43 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.00                 Driver Version: 580.00         CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA Thor                    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   40C  N/A               2W /  N/A  | Not Supported          |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|    0   N/A  N/A            3559      G   /usr/lib/xorg/Xorg                       95MiB |
|    0   N/A  N/A            3775      G   /usr/bin/gnome-shell                     58MiB |
+-----------------------------------------------------------------------------------------+
nvidia@localhost:~$ systemctl stop gdm
==== AUTHENTICATING FOR org.freedesktop.systemd1.manage-units ====
Authentication is required to stop 'gdm.service'.
Authenticating as: nvidia
Password: 
==== AUTHENTICATION COMPLETE ====
nvidia@localhost:~$ nvidia-smi -i 0
Fri Jul 24 11:58:10 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.00                 Driver Version: 580.00         CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA Thor                    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   40C  N/A               1W /  N/A  | Not Supported          |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
nvidia@localhost:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
nvidia@localhost:~$ nvidia-smi -i 0 -mig 1
Unable to enable MIG Mode for GPU 00000000:01:00.0: Insufficient Permissions
Terminating early due to previous errors.
nvidia@localhost:~$ sudo nvidia-smi -i 0 -mig 1
Enabled MIG Mode for GPU 00000000:01:00.0

Warning: persistence mode is disabled on device 00000000:01:00.0. See the Known Issues section of the nvidia-smi(1) man page for more information. Run with [--help | -h] switch to get more information on how to enable persistence mode.
All done.
nvidia@localhost:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
nvidia@localhost:~$ sudo nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
nvidia@localhost:~$ nvidia-smi
Fri Jul 24 12:00:54 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.00                 Driver Version: 580.00         CUDA Version: 13.0     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA Thor                    Off |   00000000:01:00.0 Off |                  N/A |
| N/A   40C  N/A               3W /  N/A  | Not Supported          |     N/A      Default |
|                                         |                        |              Enabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| MIG devices:                                                                            |
+------------------+----------------------------------+-----------+-----------------------+
| GPU  GI  CI  MIG |              Shared Memory-Usage |        Vol|        Shared         |
|      ID  ID  Dev |                Shared BAR1-Usage | SM     Unc| CE ENC  DEC  OFA  JPG |
|                  |                                  |        ECC|                       |
|==================+==================================+===========+=======================|
|  No MIG devices found                                                                   |
+-----------------------------------------------------------------------------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
nvidia@localhost:~$ nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|
Failed to display GPU instance profiles: Unknown Error
nvidia@localhost:~$ sudo nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|
Failed to display GPU instance profiles: Unknown Error







nvidia@localhost:~$ sudo nvidia-smi -pm 1
Enabled Legacy persistence mode for GPU 00000000:01:00.0.
All done.
nvidia@localhost:~$ sudo nvidia-smi -i 0 -mig 1
Enabled MIG Mode for GPU 00000000:01:00.0
All done.
nvidia@localhost:~$ sudo nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|
Failed to display GPU instance profiles: Unknown Error

 

driver 버전은 이런데..

NVIDIA-SMI 580.00                 Driver Version: 580.00         CUDA Version: 13.0 

[링크 : https://forums.developer.nvidia.com/t/jetson-thor-mig-enabled-but-unable-to-create-gpu-instances-driver-580-00-l4t-r38-2/372073]

 

jetpack 버전은 38.4 같다. 일단 위에서 JetPack 7.2GA/R39.2 라고 했으니 업데이드 해봐야하나?

$ dpkg-query --show nvidia-l4t-core
nvidia-l4t-core 38.4.0-20251230160601

[링크 : https://yoo-hk.tistory.com/555]

 

 

--

싹다 밀어버리고 컴백

$ nvidia-smi
Fri Jul 24 17:02:29 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.78                 Driver Version: 595.78         CUDA Version: 13.2     |

$ dpkg-query --show nvidia-l4t-core
nvidia-l4t-core 39.2.0-20260601141651

 

일단 잡을 녀석들 죽이고

$ sudo systemctl stop gdm
$ sudo systemctl stop nvargus-daemon

 

일단 persistent mode 하지 않고 하면 목록에서 에러는 나지 않는데 실패

nvidia@nvidia:~$ nvidia-smi -i 0 -mig 1
Unable to enable MIG Mode for GPU 00000000:01:00.0: Insufficient Permissions
Terminating early due to previous errors.
nvidia@nvidia:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
nvidia@nvidia:~$ nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|
nvidia@nvidia:~$ sudo nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|
nvidia@nvidia:~$ nvidia-smi mig

    mig -- Multi Instance GPU management.

    Usage: nvidia-smi mig [options]

    Options include:
    [-h | --help]: Display help information.
    [-i | --id]: Enumeration index, PCI bus ID or UUID.
                 Provide comma separated values for more than one device.
    [-gi | --gpu-instance-id]: GPU instance ID.
                               Provide comma separated values for more than one GPU instance.
    [-ci | --compute-instance-id]: Compute instance ID.
                                   Provide comma separated values for more than one compute 
                                   instance.
    [-lgip | --list-gpu-instance-profiles]: List supported GPU instance profiles.
                                            Option -i can be used to restrict the command to
                                            run on a specific GPU.
    [-lgipp | --list-gpu-instance-possible-placements]: List possible GPU instance placements
                                                        in the following format, {Start}:Size.
                                                        Option -i can be used to restrict the
                                                        command to run on a specific GPU.
    [-C | --default-compute-instance]: Create compute instance with the default profile when used
                                       with the option to create a GPU instance (-cgi).
    [-cgi | --create-gpu-instance]: Create GPU instances for the given profile tuples. A profile
                                    tuple consists of a profile name or ID and an optional placement
                                    specifier, which consists of a colon and a placement start index.
                                    Provide comma separated values for more than one profile tuple.
                                    Option -i can be used to restrict the command to run on
                                    a specific GPU.
    [-dgi | --destroy-gpu-instance]: Destroy GPU instances.
                                     Options -i and -gi can be used individually or combined
                                     to restrict the command to run on a specific GPU or GPU
                                     instance.
    [-lgi | --list-gpu-instances]: List GPU instances.
                                   Option -i can be used to restrict the command to run on a
                                   specific GPU.
    [-lcip | --list-compute-instance-profiles]: List supported compute instance profiles.
                                                Options -i and -gi can be used individually or
                                                combined to restrict the command to run on a
                                                specific GPU or GPU instance.
    [-lcipp | --list-compute-instance-possible-placements]: List possible compute instance placements
                                                            in the following format, {Start}:Size.
                                                            Options -i and -gi can be used individually or
                                                            combined to restrict the command to run on a
                                                            specific GPU or GPU instance.
    [-cci | --create-compute-instance]: Create compute instance for the given profile name or IDs.
                                        Provide comma separated values for more than one profile.
                                        If no profile name or ID is given, then the default*
                                        compute instance profile ID will be used. Options -i and
                                        -gi can be used individually or combined to restrict the
                                        command to run on a specific GPU or GPU instance.
    [-dci | --destroy-compute-instance]: Destroy compute instances.
                                         Options -i, -gi and -ci can be used individually or
                                         combined to restrict the command to run on a specific
                                         GPU or GPU instance or compute instance.
    [-lci | --list-compute-instances]: List compute instances.
                                       Options -i and -gi can be used individually or combined
                                       to restrict the command to run on a specific GPU or GPU
                                       instance.

 

persistent mode를 활성화 해주어야 한다. 물론 persistent mode 한다고 해서

전원 off 하고 다시 켠다고 또 되는건 아닌듯. 그럼 왜 persistent mode라고 하는겨?

nvidia@nvidia:~$ sudo nvidia-smi -pm 1
Enabled Legacy persistence mode for GPU 00000000:01:00.0.
All done.
nvidia@nvidia:~$ sudo nvidia-smi -i 0 -mig 1
Enabled MIG Mode for GPU 00000000:01:00.0
All done.
nvidia@nvidia:~$ sudo nvidia-smi mig -lgip
+-------------------------------------------------------------------------------+
| GPU instance profiles:                                                        |
| GPU   Name               ID    Instances   Memory     P2P    SM    DEC   ENC  |
|                                Free/Total   GiB              CE    JPEG  OFA  |
|===============================================================================|
|   0  MIG 1g.0gb+me       78     1/1        0.00       No      6     1     1   |
|                                                               1     1     1   |
+-------------------------------------------------------------------------------+
|   0  MIG 1g.0gb          80     1/1        0.00       No      6     1     1   |
|                                                               1     1     0   |
+-------------------------------------------------------------------------------+
|   0  MIG 1g.0gb+gfx      81     1/1        0.00       No      6     1     1   |
|                                                               1     1     0   |
+-------------------------------------------------------------------------------+
|   0  MIG 2g.0gb          82     1/1        0.00       No     12     1     1   |
|                                                               1     1     0   |
+-------------------------------------------------------------------------------+
|   0  MIG 2g.0gb+gfx      83     1/1        0.00       No     12     1     1   |
|                                                               1     1     0   |
+-------------------------------------------------------------------------------+
|   0  MIG 3g.0gb           0     1/1        0.00       No     20     2     2   |
|                                                               1     2     1   |
+-------------------------------------------------------------------------------+
|   0  MIG 3g.0gb+gfx      32     1/1        0.00       No     20     2     2   |
|                                                               1     2     1   |
+-------------------------------------------------------------------------------+

 

아무튼 이렇게 해야지 리셋도 된다.

nvidia@nvidia:~$ sudo systemctl status nvargus-daemon
● nvargus-daemon.service - Argus daemon
     Loaded: loaded (/etc/systemd/system/nvargus-daemon.service; enabled; preset: enabled)
     Active: active (running) since Fri 2026-07-24 17:00:50 KST; 14min ago
   Main PID: 2555 (nvargus-daemon)
      Tasks: 3 (limit: 150257)
     Memory: 68.8M (peak: 108.8M)
        CPU: 285ms
     CGroup: /system.slice/nvargus-daemon.service
             └─2555 /usr/sbin/nvargus-daemon

Jul 24 17:00:54 nvidia nvargus-daemon[2555]: CaptureService:stop: is requested --------------------------
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: SCF: Error 0x00000008: Service not running (in src/services/power/PowerService.cpp, function stopS>
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: SCF: Error 0x00000008:  (propagating from src/components/ServiceHost.cpp, function stopServicesInt>
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: SCF: Error 0x0000000f:  (propagating from src/api/CameraDriver.cpp, function initialize(), line 18>
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: SCF: Error 0x00000008: Services are already stopped (in src/components/ServiceHost.cpp, function s>
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: SCF: Error 0x0000000f:  (propagating from src/api/CameraDriver.cpp, function getCameraDriver(), li>
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: (Argus) Error 0x0000000f:  (propagating from src/api/GlobalProcessState.cpp, function createCamera>
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: === gst-plugin-scanner2811)[2811]: CameraProvider failed to initialize
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: === gst-plugin-scanner2811)[2811]: Connection closed (FFFF949B88C0)
Jul 24 17:00:55 nvidia nvargus-daemon[2555]: === gst-plugin-scanner2811)[2811]: Connection cleaned up (FFFF949B88C0)
nvidia@nvidia:~$ sudo systemctl stop nvargus-daemon
nvidia@nvidia:~$ sudo nvidia-smi --gpu-reset -i 0
GPU 00000000:01:00.0 was successfully reset.
All done.

 

일단 다 죽이고

$ sudo nvidia-smi mig -dci && sudo nvidia-smi mig -dgi
No GPU instances found: Not Found

 

config 1 - 0
config 2 - 32
config 3 - 78, 83
config 4 - 80, 83
config 5 - 81, 82
config 6 - 82
config 7 - 83
config 8 - 78
config 9 - 80
config 10- 81

 

$ sudo nvidia-smi mig -cgi 80,83
Successfully created GPU instance ID  2 on GPU  0 using profile MIG 1g.0gb (ID 80)
Successfully created GPU instance ID  1 on GPU  0 using profile MIG 2g.0gb+gfx (ID 83)
nvidia@nvidia:~$ nvidia-smi
Fri Jul 24 17:50:43 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.78                 Driver Version: 595.78         CUDA Version: 13.2     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA Thor                    On  |   00000000:01:00.0 Off |                  N/A |
| N/A   42C  N/A               4W /  N/A  | Not Supported          |     N/A      Default |
|                                         |                        |              Enabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| MIG devices:                                                                            |
+------------------+----------------------------------+-----------+-----------------------+
| GPU  GI  CI  MIG |              Shared Memory-Usage |        Vol|        Shared         |
|      ID  ID  Dev |                Shared BAR1-Usage | SM     Unc| CE ENC  DEC  OFA  JPG |
|                  |                                  |        ECC|                       |
|==================+==================================+===========+=======================|
|  No MIG devices found                                                                   |
+-----------------------------------------------------------------------------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

 

nvidia@nvidia:~$ sudo nvidia-smi mig -cgi 80,83 -C
Successfully created GPU instance ID  2 on GPU  0 using profile MIG 1g.0gb (ID 80)
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  2 using profile MIG 1g.0gb (ID  0)
Successfully created GPU instance ID  1 on GPU  0 using profile MIG 2g.0gb+gfx (ID 83)
Successfully created compute instance ID  0 on GPU  0 GPU instance ID  1 using profile MIG 2g.0gb (ID  1)
nvidia@nvidia:~$ nvidia-smi
Fri Jul 24 17:53:35 2026       
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.78                 Driver Version: 595.78         CUDA Version: 13.2     |
+-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA Thor                    On  |   00000000:01:00.0 Off |                  N/A |
| N/A   41C  N/A               4W /  N/A  | Not Supported          |     N/A      Default |
|                                         |                        |              Enabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| MIG devices:                                                                            |
+------------------+----------------------------------+-----------+-----------------------+
| GPU  GI  CI  MIG |              Shared Memory-Usage |        Vol|        Shared         |
|      ID  ID  Dev |                Shared BAR1-Usage | SM     Unc| CE ENC  DEC  OFA  JPG |
|                  |                                  |        ECC|                       |
|==================+==================================+===========+=======================|
|  0    1   0   0  |                 N/A              | 12      0 |  1   1    1    0    1 |
|                  |                 N/A              |           |                       |
+------------------+----------------------------------+-----------+-----------------------+
|  0    2   0   1  |                 N/A              |  6      0 |  1   1    1    0    1 |
|                  |                 N/A              |           |                       |
+------------------+----------------------------------+-----------+-----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+
nvidia@nvidia:~$ nvidia-smi -L
GPU 0: NVIDIA Thor (UUID: GPU-a7c66ad2-6dbb-0ab8-c1a2-37ba6dba3600)
  MIG 2g.0gb      Device  0: (UUID: MIG-031bd8db-02a6-5420-968d-7fcdf907f7e4)
  MIG 1g.0gb      Device  1: (UUID: MIG-60591740-f26d-525b-ad5a-5f5ba115bc71)

 

 

+

2026.07.25

[링크 : https://toss.tech/article/toss-securities-gpu-mig]

저작자표시 (새창열림)

'embeded > jetson' 카테고리의 다른 글

nvidia-smi mig  (0) 2026.07.26
nvidia jetson agx thor dev kit 설치 - USB boot  (0) 2026.07.24
nvidia jetson 아니고 nvidia thor  (0) 2026.07.24
jetson nano + llama.cpp vulkan  (0) 2026.07.21
jetson nano jtop  (0) 2026.07.20
Posted by 구차니

티스토리툴바