Using Arc Battlemage on a Dell poweredge r730

[Rileyk]

Hello,

I am looking to buy an Arc B580 to put in my Dell PowerEdge r730 which has 2x Xeon E5 2683v4 CPUs. I have some questions regarding driver compatibility for the card given the age of the system and lack of native ReBAR support in the BIOS. 

I only intend to use the card for media transcode and compute applications on a non-graphical Linux environment, specifically Alpine Linux.

The system does support 'Above 4G Decode', which I have enabled however, I would need to apply a BIOS mod to support ReBAR which I do not want to do as it is annoying to get physical access to the EEPROM and to test if the mod worked. From what little I understand both of these features have a similar effect of allowing a larger mapping of device addresses to be accessible. 

Getting to the question, would the card work without applying the above-mentioned bios mod for my intended applications in the specified environment? I understand this is an unusual use case and I do not expect to be given any guarantees however I would still value the advice. 

As a bonus question, does Intel have any considerations for removing ReBAR as a requirement for Arc Battlemage? Given Arc appears to be a more budget-oriented lineup it would make sense to remove said requirement if it is technically possible.

Kind regards,

Riley

[JeanetteC_Intel]

Hello Rileyk,

Thank you for posting in Intel Communities.

The Intel® Arc™ B580 Graphics may not run as expected since your Dell PowerEdge R730 does not conform to the supported hardware listed on the Intel® Arc™ B-Series Graphics – Desktop Quick Start Guide. For you to have the GPU running within public specifications for Intel Arc 580 Kindly check the system requirements (particularly the supported CPU and OS) on the link below.

Enabling ReBar is constant event from the start of ARC Graphics was introduced (from A series), but I can still check internally if removing ReBAR as a requirement will be possible in the future.

I'll post back when I have updates.

Best regards,

JeanetteC.

Intel® Customer Support Technician

[JeanetteC_Intel · Accepted solution]

Hello Rileyk,

Resizable BAR (ReBAR) is a feature that allows the CPU to access the entire GPU memory simultaneously, potentially boosting performance in certain situations. However, for tasks like media transcoding and computing, which are not gaming-related, the benefits of ReBAR may not be as pronounced.

Since your system supports and has enabled Above 4G Decoding, this should facilitate the handling of larger memory spaces, which is advantageous for modern GPUs.

While Intel's Arc GPUs are optimized for modern systems, they can still operate in older systems, albeit with some limitations. The absence of ReBAR might not prevent the card from functioning, but it could slightly restrict performance in specific scenarios. For your use case, which doesn't involve gaming, this limitation is likely not a major concern.

For Linux compatibility, it's crucial to have the latest drivers and kernel updates installed. You may need to compile drivers or adjust kernel parameters to achieve optimal performance.

Regarding the potential removal of ReBAR as a requirement, Intel has not made any announcements. ReBAR can enhance performance, particularly in gaming, but its importance can vary based on the application.

In conclusion, the Arc B580 should be suitable for your media transcoding and compute tasks, even without ReBAR, though you might not achieve maximum performance. Testing the card in your specific environment will be the best way to assess its effectiveness.

For more detailed information on system requirements for Intel Arc graphics, please refer to the official system requirements article: Intel® Arc™ Graphics – Desktop Quick Start Guide

I hope this answer your questions.

Best regards,

JeanetteC.

Intel Customer Support Technician

[JeanetteC_Intel]

 Hello Rileyk,

Were you able to check my previous post?

Let me know if you have any questions. 

Best regards,

JeanetteC.

Intel Customer Support Technician

[JeanetteC_Intel]

Hello Rileyk,

Since I have not seen any new post/reply, I will now proceed in closing this thread. Should you need assistance in the future, please submit a new question as this thread will no longer be monitored.

Sincerely,

JeanetteC.

Intel Customer Support Technician

[Rileyk]

Sorry for the extremely delayed response,

I ended up buying a Sparkle Arc Pro B60 and installed it into my R730 and it works fine. I am able to run media transcoding, Vulkan compute (tested via llama.cpp), OpenCL kernels and pytorch through a Ubuntu docker container. 

Some things of note that I have encountered are;

 - The GPU max power is reported as 440 W in nvtop which is 2x what it should be and is definitely unattainable.

 - The maximum observed power consumed is 110 W in nvtop hinting under utilization of the compute cores.

 - Pytorch / onednn does not seem to support offloading GPU memory to host memory automatically. This is not as big of a deal for me personally due to the GPU only having PCIe 3.0 x8 of PCIe bandwidth which would cripple performance. If you encounter an OOM on the GPU, I have found that you need to restart the Jupyter kernel to free the memory.

 - I have had some strange onednn crashes when training pytorch convolutional models, typically after ~40 mins. Maybe a semi-handled error in onednn where it would print error info but crash/exit before returning to python with an exception? That being said it may have been fixed as I have not had any issues for the last month.

 - (seeming) Lack of over/under clocking/volting support or power limit controls. Given the lower-than-possible power usage, I suspect memory bandwidth is not fast enough to keep up with the XMX and Xe cores and overclocking the memory would be useful to boost it.

 - Idle power consumption of the GPU is always 30 W and does not seem possible to lower it further short of disabling the card entirely. Probably a skill issue on my part with configuring Linux.

Kind regards,

Riley