Nvidia Ada card RAM not detected correctly

Message boards : GPUs : Nvidia Ada card RAM not detected correctly
Message board moderation

To post messages, you must log in.

AuthorMessage
ED Labs
New member

Send message
Joined: 19 Jul 24
Posts: 4
Italy
Message 114287 - Posted: 19 Jul 2024, 6:43:58 UTC

Hello,
FYI my 20GB Nvidia RTX 4000 SFF GPU appears to be detected by BOINC as only having 4GB of total video RAM.

The GPU Grid website reports 4GB
NVIDIA NVIDIA RTX 4000 SFF Ada Generation (4095MB) driver: 555.42

BOINC client config files also report 4GB
   <name>NVIDIA RTX 4000 SFF Ada Generation</name>
   <available_ram>4152360960.000000</available_ram>

BOINC logs report 4GB with CUDA but 20GB with OpenCL:
Thu 18 Jul 2024 06:13:00 PM CEST |  | CUDA: NVIDIA GPU 0: NVIDIA RTX 4000 SFF Ada Generation (driver version 555.42, CUDA version 12.5, compute capability 8.9, 4096MB, 3960MB available, 19169 GFLOPS peak)
Thu 18 Jul 2024 06:13:00 PM CEST |  | OpenCL: NVIDIA GPU 0: NVIDIA RTX 4000 SFF Ada Generation (driver version 555.42.06, device version OpenCL 3.0 CUDA, 20055MB, 3960MB available, 19169 GFLOPS peak)

A seemingly related bug from 2017 appears to have been "solved" in 2022.
https://github.com/BOINC/boinc/issues/1773

In the last few days GPU Grid used between 4 and 8GB of video RAM. Right now it's using just 1,5GB and 100% computation. The card is currently not connected to any displays nor running other processes (save for a 4MB Xorg one), so almost all of its resources are available for BOINC computation.
ID: 114287 · Report as offensive     Reply Quote
Dr Who Fan
Avatar

Send message
Joined: 10 May 07
Posts: 1378
United States
Message 114293 - Posted: 19 Jul 2024, 21:27:22 UTC - in response to Message 114287.  

You don't say what version of BOINC you are using?
Yes, it was fixed in version 7.20.0 and higher.
So if you are using any version prior to that, download and install the latest version for you operating system.
https://boinc.berkeley.edu/download_all.php
ID: 114293 · Report as offensive     Reply Quote
ED Labs
New member

Send message
Joined: 19 Jul 24
Posts: 4
Italy
Message 114294 - Posted: 19 Jul 2024, 23:09:19 UTC - in response to Message 114293.  

Thanks for the prompt reply, you're right.

I was unknowingly using an outdated version of BOINC. I thought I was using the latest version because I had just installed BOINC from the repository of the latest Linux Mint (21.3 "Virginia"). The latest version of Mint is based upon an old version of Ubuntu (22.04 "Jammy") still maintained, whose repository contains BOINC 7.18.1 from 2021 (before the aforementioned bug was fixed). Instead the latest version of Ubuntu (24.04 "Noble") contains BOINC 7.24.1 without the Nvidia RAM bug.

I will consider reinstalling BOINC from the dedicated PPA mentioned in the official BOINC page. On a related note though it would be useful if BOINC reported the usage of an outdated version in its "notices" tab.

Reference: https://boinc.berkeley.edu/wiki/Installing_BOINC_on_Ubuntu
ID: 114294 · Report as offensive     Reply Quote
ED Labs
New member

Send message
Joined: 19 Jul 24
Posts: 4
Italy
Message 114295 - Posted: 19 Jul 2024, 23:11:30 UTC - in response to Message 114294.  

Correction: I will reinstall using the link you kindly provided.
ID: 114295 · Report as offensive     Reply Quote
ED Labs
New member

Send message
Joined: 19 Jul 24
Posts: 4
Italy
Message 114296 - Posted: 20 Jul 2024, 0:48:07 UTC - in response to Message 114295.  

I reinstalled from the official Berkeley repository and after some troubleshooting everything worked. BOINC now recognizes all video RAM. I leave here the troubleshooting I did in case it might help others.

After reinstalling, BOINC Manager would not start with error "Invalid client RPC password. Try reinstalling BOINC". An ordinary reinstallation did not help, nor did following instructions in the first few online results debating the same issue. What worked for me was purging packages, removing the boinc group and user, then removing from /etc and ~/ all related directories and configuration files having names or contents related with BOINC.

There might be a typo on the page with the installation instructions at https://boinc.berkeley.edu/linux_install.php?os_num=5&build=stable
The suggested apt-add-repository command line does enclose the single argument in quotes, which was required for it to work on my system
sudo apt-add-repository 'deb https://boinc.berkeley.edu/dl/linux/stable/jammy jammy main'
ID: 114296 · Report as offensive     Reply Quote

Message boards : GPUs : Nvidia Ada card RAM not detected correctly

Copyright © 2024 University of California.
Permission is granted to copy, distribute and/or modify this document under the terms of the GNU Free Documentation License, Version 1.2 or any later version published by the Free Software Foundation.