Using accel-sim with SM 8.6

144 views
Skip to first unread message

Shougang Yuan

<syuan3@ncsu.edu>
unread,
Jan 3, 2022, 11:41:07 PM1/3/22
to accel-sim
Hi, All,

I am recently trying to transfer from GPGPU-Sim to Accel-Sim, and my local host has RTX 3060 Ti with SM compute capability of 8.6. I got some error message--"NVBit ERROR: SM 8.6 name NVIDIA GeForce RTX 3060 Ti not supported".

I checked the NVBit git repo, and found that it should support SM 8.6. Is there any trick for Accel-Sim with SM 8.6?

Best regards
Shougang

Junrui Pan

<panjunrui100@gmail.com>
unread,
Jan 4, 2022, 8:52:19 PM1/4/22
to accel-sim
Hi, 

According to nvbit repo, SM86 support was added in v1.5.3. Accel-Sim currently uses nvbit v1.5. 
If you want you may update the scripts in the util folder to use the latest nvbit for the tracer. 

Thanks,
Junrui

Shougang Yuan

<syuan3@ncsu.edu>
unread,
Jan 5, 2022, 11:09:45 PM1/5/22
to Junrui Pan, accel-sim
I already replaced the nvibit with version 1.5.4, however, I am still getting the same error, and can not get the trace for the benchmarks

--
You received this message because you are subscribed to the Google Groups "accel-sim" group.
To unsubscribe from this group and stop receiving emails from it, send an email to accel-sim+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/accel-sim/8e3e70ef-5fc0-474a-a780-292220e89186n%40googlegroups.com.

Mahmood

<mahmood.nt@gmail.com>
unread,
Jan 6, 2022, 5:51:32 AM1/6/22
to Shougang Yuan, Junrui Pan, accel-sim

Hi,

Did you use the dev version?

 

Regards,

Mahmood

 

Sent from Mail for Windows

Shougang Yuan

<syuan3@ncsu.edu>
unread,
Jan 6, 2022, 1:33:19 PM1/6/22
to Mahmood, Junrui Pan, accel-sim
Yes, I am using the dev branch, my cuda toolkit is 11.0, my host has RTX3060 Ti.

Best regards.
Shougang

Pan Junrui

<panjunrui100@gmail.com>
unread,
Jan 6, 2022, 1:42:41 PM1/6/22
to Shougang Yuan, Mahmood, accel-sim
Not 100 percent sure about this. But I do remember you need at least CUDA 11.1 for native support on 30s cards. Have you tried a newer CUDA?

Thanks,
Junrui

From: Shougang Yuan <syu...@ncsu.edu>
Sent: Thursday, January 6, 2022 1:33:07 PM
To: Mahmood <mahmo...@gmail.com>
Cc: Junrui Pan <panjun...@gmail.com>; accel-sim <acce...@googlegroups.com>

Shougang Yuan

<syuan3@ncsu.edu>
unread,
Jan 6, 2022, 2:02:08 PM1/6/22
to Pan Junrui, Mahmood, accel-sim
I have CUDA 7.0 to CUDA 11.4 all on my machine, and I have tried CUDA 10, CUDA 11, CUDA 11.2, I still get the same error message. I will try on another machine and see what happens.

Thanks and regards.
Shougang

Mahmood Naderan

<mahmood.nt@gmail.com>
unread,
Jan 6, 2022, 2:04:43 PM1/6/22
to Shougang Yuan, Pan Junrui, accel-sim
Are you able to run the examples, e.g. mem_trace with your application? If not, then you should ask nvbit.

Regards,
Mahmood

Pan Junrui

<panjunrui100@gmail.com>
unread,
Jan 6, 2022, 2:25:39 PM1/6/22
to Mahmood Naderan, Shougang Yuan, accel-sim
Yes it looks like a nvbit problem. But we were able to generate traces for a RTX3070 on our machine with CUDA11.1.

Please make sure to clean the tracer before running make with different CUDA version.


If this still doesn't work, you may use the traces we provided to start with. They are generated with a 3070.

Thanks,
Junrui
From: Mahmood Naderan <mahmo...@gmail.com>
Sent: Thursday, January 6, 2022 2:04:29 PM
To: Shougang Yuan <syu...@ncsu.edu>
Cc: Pan Junrui <panjun...@gmail.com>; accel-sim <acce...@googlegroups.com>
Reply all
Reply to author
Forward
0 new messages