Dear Nvidia Support Team,
From the overview page of Aerial CUDA-Accelerated RAN container, there are prerequisites as below,
Does it mean that Aerial CUDA-Accelerated RAN only be supported with GPU lists above?
If that, is there any plan to extend supporting GPU lists?
Thank you for reaching out with the queries. The listed converged accelerator (AX800, A100X) and the GPU+NIC combinations (A100+CX6DX, GH MGX+BF3) are the CUDA capable GPU with PTP enabled Nvidia NIC card we have tested Aerial with. You are free to try other CUDA capable GPUs but we have not tested with them. Please let us know what GPU you are interested in exercising.
Thanks
Thanks for your response,
We are in the process of setting up a test environment, and it includes tests for Aerial as well.
Currently, we are reviewing the specifications required for setting up the environment, and I will provide more details once they are finalized.
If there are any recommended minimum system requirements for Aerial, please let me know.
It would be very helpful to me.
Best regards
@jixu do you think it’s possible to run it using a CX6-LX NIC + A100 GPU? Or the CX6-DX is mandatory?
@emelao,
CX6-LX doesn’t list PTP/PHC2Sys support, so CX6DX is needed for the setup with A100 GPU.
Really? When I run ethtool -T <interface> I get:
Time stamping parameters for <interface>:
Capabilities:
hardware-transmit
hardware-receive
hardware-raw-clock
PTP Hardware Clock: 4
Hardware Transmit Timestamp Modes:
off
on
Hardware Receive Filter Modes:
none
all
Or do I need software support as well? Because here the PTP service and the PHC2Sys is running without problems:
systemctl status ptp4l.service
● ptp4l.service - Precision Time Protocol (PTP) service
Loaded: loaded (/lib/systemd/system/ptp4l.service; enabled; vendor preset: enabled)
Active: active (running) since Tue 2024-05-07 10:43:05 -03; 1 day 3h ago
Docs: man:ptp4l
Main PID: 3102 (ptp4l)
Tasks: 1 (limit: 78127)
Memory: 6.0M
CPU: 1min 43.835s
CGroup: /system.slice/ptp4l.service
└─3102 /usr/sbin/ptp4l -f /etc/ptp.conf
May 08 14:18:38 supermicro taskset[3102]: ptp4l[99475.363]: rms 4 max 8 freq -103014 +/- 7 delay 205 +/- 1
May 08 14:18:39 supermicro taskset[3102]: ptp4l[99476.488]: rms 5 max 10 freq -103016 +/- 8 delay 205 +/- 0
May 08 14:18:40 supermicro taskset[3102]: ptp4l[99477.613]: rms 2 max 6 freq -103015 +/- 4 delay 206 +/- 1
May 08 14:18:41 supermicro taskset[3102]: ptp4l[99478.738]: rms 4 max 6 freq -103010 +/- 5 delay 206 +/- 1
May 08 14:18:42 supermicro taskset[3102]: ptp4l[99479.863]: rms 4 max 6 freq -103015 +/- 6 delay 206 +/- 0
May 08 14:18:43 supermicro taskset[3102]: ptp4l[99480.988]: rms 3 max 9 freq -103013 +/- 6 delay 205 +/- 0
May 08 14:18:45 supermicro taskset[3102]: ptp4l[99482.113]: rms 3 max 7 freq -103012 +/- 6 delay 204 +/- 1
May 08 14:18:46 supermicro taskset[3102]: ptp4l[99483.237]: rms 4 max 10 freq -103018 +/- 7 delay 205 +/- 1
May 08 14:18:47 supermicro taskset[3102]: ptp4l[99484.362]: rms 6 max 10 freq -103004 +/- 7 delay 204 +/- 0
May 08 14:18:48 supermicro taskset[3102]: ptp4l[99485.487]: rms 4 max 9 freq -103012 +/- 6 delay 205 +/- 0
systemctl status phc2sys.service
● phc2sys.service - Synchronize system clock or PTP hardware clock (PHC)
Loaded: loaded (/lib/systemd/system/phc2sys.service; enabled; vendor preset: enabled)
Active: active (running) since Tue 2024-05-07 13:45:00 -03; 24h ago
Docs: man:phc2sys
Main PID: 43647 (sudo)
Tasks: 2 (limit: 78127)
Memory: 4.5M
CPU: 6min 23.734s
CGroup: /system.slice/phc2sys.service
├─43647 sudo /usr/sbin/phc2sys -s /dev/ptp4 -c CLOCK_REALTIME -n 24 -O 0 -R 256 -u 256
└─43656 /usr/sbin/phc2sys -s /dev/ptp4 -c CLOCK_REALTIME -n 24 -O 0 -R 256 -u 256
May 08 14:19:00 supermicro phc2sys[43656]: [99497.335] CLOCK_REALTIME rms 11 max 30 freq +2564 +/- 80 delay 507 +/- 20
May 08 14:19:01 supermicro phc2sys[43656]: [99498.353] CLOCK_REALTIME rms 8 max 21 freq +2576 +/- 51 delay 510 +/- 19
May 08 14:19:02 supermicro phc2sys[43656]: [99499.374] CLOCK_REALTIME rms 6 max 17 freq +2577 +/- 18 delay 510 +/- 19
May 08 14:19:03 supermicro phc2sys[43656]: [99500.391] CLOCK_REALTIME rms 8 max 23 freq +2573 +/- 45 delay 507 +/- 20
May 08 14:19:04 supermicro phc2sys[43656]: [99501.409] CLOCK_REALTIME rms 8 max 18 freq +2565 +/- 59 delay 511 +/- 18
May 08 14:19:05 supermicro phc2sys[43656]: [99502.426] CLOCK_REALTIME rms 7 max 19 freq +2590 +/- 39 delay 509 +/- 19
May 08 14:19:06 supermicro phc2sys[43656]: [99503.446] CLOCK_REALTIME rms 10 max 24 freq +2571 +/- 72 delay 508 +/- 20
May 08 14:19:07 supermicro phc2sys[43656]: [99504.463] CLOCK_REALTIME rms 10 max 30 freq +2567 +/- 86 delay 506 +/- 20
May 08 14:19:08 supermicro phc2sys[43656]: [99505.481] CLOCK_REALTIME rms 8 max 22 freq +2591 +/- 49 delay 512 +/- 18
May 08 14:19:09 supermicro phc2sys[43656]: [99506.498] CLOCK_REALTIME rms 8 max 25 freq +2570 +/- 34 delay 510 +/- 18
Hi tojsm,
We recommended the systems below for running Aerial,
Gigabyte x86+A100+CX6DX
Dell R750+A100X
Dell R750+AX800
MGX +BF3
Here is the installation guide for the qualified platform and component: cuBB Installation Guide - NVIDIA Docs
Hi emelao,
The CX6-LX does’t have ‘Nanosecond-level accuracy’ that is available on CX6DX.
Hi,
Could you tell me where I can buy an AX800 or A100X?
@alonsoch Thank you very much for being interested in Aerial CUDA-accelerated RAN and reaching out with the query. A100X can perhaps be obtained via resellers. Arrow may have some. AX800 is not available for MP yet, only QS.
Hi,
Does A30X support Aerial CUDA-Accelerated RAN?
Hi：
Does A30X support Aerial CUDA-Accelerated RAN?
We purchased the A30X in 2023, and an Aerial CUDA-Accelerated RAN has not yet been released
Hi @lyy2,
Welcome to the Aerial Forum!
A30X is unfortunately not supported by Aerial CUDA-Accelerated RAN.
Aerial CUDA-Accelerated RAN can be run with A100X, A800X or GH200.
Thank you.
Hello bkecicioglu, does Aerial RAN supports L40S? Unfortunately, the H100 is beyond our budget and A100 is not available in Taiwan. Our distributor has suggested the L40S as an alternative. Thank you for your assistance.
Hi @junxian,
We have not tested and qualified Aerial with L40S. If you decide to use L40S and have any issues, please let us know.
We will be happy to help.
Thank you.
@bkecicioglu Thank you.
We have successfully installed the cuBB container on our A6000 GPU and have run several examples of cuPHY and cuMAC using the provided testing vectors. Our next step is to build a server and connect it to a real O-RU.
Given that the tests for cuMAC and cuPHY have passed, can we conclude that the A6000 meets the necessary performance requirements for these components? Additionally, are there any further tests you would recommend before we proceed with system integration?
Thank you for your support.
Hi @junxian
which NIC you used in your setup?
can you share the details
Thanks
We initially used a CX6-DX NIC for testing, but it did not work as expected. This might be due to the CX6-DX + A100 being part of an older framework, and the A100 is no longer available for procurement. As the next step, we plan to test using a BF3 NIC.