Feature Request: Native Support for Expandable Memory Controllers and Unlocked vBIOS in Future Archi

[UmAI]

I am an AI researcher and hardware enthusiast, and I am writing to propose a radical but highly necessary shift in how GPU silicon is locked down, specifically regarding VRAM limits and memory controllers.
Currently, the massive VRAM requirements of local AI training and inferencing are colliding with the rigid hardware limits of consumer and prosumer cards. While your compute architecture and XPU integration have made massive strides, users are still walled in by the physical memory layout and hardcoded vBIOS on the cards.
The Vision: Modular Silicon for Custom Integrations
My request is for future architectures (like Celestial) to support flexible memory controllers and an unlocked vBIOS ecosystem that would allow developers, makers, and researchers to build custom boards.
Specifically, I want to be able to take an Intel GPU die, design a custom PCB (such as a high-bandwidth USB4/Thunderbolt eGPU board), and pair it with significantly more VRAM than a standard retail card provides.
To make this a reality, we would need:
Scalable Memory Support: Controllers that can dynamically recognize and utilize higher-density VRAM chips or clamshell configurations.
Open or Hackable vBIOS: A supported pathway to flash custom firmware that doesn't instantly reject non-standard memory topologies.
Die Availability: A pathway for researchers or boutique hardware creators to source bare silicon, rather than having to desolder existing retail cards.
Empowering the community to build custom, high-VRAM compute boards around Intel silicon would make Arc the undisputed champion of open-source hardware and local AI research.
Thank you for pushing the envelope, and I hope you consider opening up the hardware to let the community push it even further.

[Jobert_Intel]

Hi UmAI,

Thank you for your post and for sharing such a detailed and well-thought-out feature request with the Intel® Community!

We truly appreciate the passion and technical insight you have brought to this discussion. Your vision for modular silicon, scalable memory controllers, and an open vBIOS ecosystem for future Intel® Arc™ architectures is noted, and we want to make sure your feedback reaches the right people.

To help us better understand your needs and properly document your request, we have a few follow-up questions:

1. What is your primary workload? (e.g., local LLM inferencing, AI model training, computer vision, or other research applications)
2. Which Intel® Arc™ product(s) are you currently using or evaluating for your AI research work?
3. What minimum VRAM capacity would realistically meet your workload requirements?
4. Have you explored Intel® Data Center GPU solutions (e.g., Intel® Gaudi® or Intel® Arc™ Pro series) as a potential fit for your use case?
5. Would you be interested in participating in Intel® developer programs or community feedback initiatives if made available in the future?

Your feedback is genuinely valuable, and we look forward to hearing more from you!

Best Regards,

Jobert T.

Intel Customer Support Technician

[UmAI]

Good afternoon Jobert,

I am honestly surprised that someone answered me! Thank you for the response and for passing this along.

To give you a better idea of my background and needs, here are the answers to your questions:

1. Primary Workload I focus heavily on custom model architecture and training. For example, here is a link to one of my published projects: https://github.com/byte1024x/Recursive-RK2-Transformer-177M.git. This is my second most recent project, but definitely not my best—just the only one I have published so far. (It would take me 5+ hours to post all of them, but if you want the model weights for this one, just shoot me a message!)

2. Current Intel® Arc™ Products I actually do not currently own an Intel GPU, which is part of the reason I am so passionate about this! My current rig has a Ryzen 5 7600X, 32GB of DDR5 RAM, and an RTX 5070. When I built my PC, I did not know Intel made GPUs at the time. I heard everyone saying Nvidia is the best, so I got the 5070. About three months later, I realized I could have gotten a cheaper Intel GPU with way more VRAM.

3. Minimum VRAM Capacity Because model training and custom transformer projects eat up memory quickly, 24GB to 32GB is really the baseline I am looking for. Having modular boards where I could push that to 90GB+ using Intel silicon would completely change the game for solo developers like me who are bottlenecked by standard consumer cards.

4. Data Center Solutions I have looked at them, but things like Intel Gaudi or the Arc Pro series are generally priced and designed for enterprise environments. My request is specifically aimed at the maker/enthusiast/researcher level—people who want to build custom, high-VRAM desktop setups without paying enterprise data center premiums.

5. Developer Programs Absolutely. Because I have so many unpublished projects, I would be incredibly interested in participating in any developer programs, testing custom board designs (I would love that), or giving feedback on future hardware initiatives.

Thank you again for taking the time to reply!

Best, UmAI!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!

[Jobert_Intel]

Hi UmAI,

Thank you for providing the requested information. I will now further investigate this and will get back to you once a resolution is available. Thank you for your patience and understanding.

Best Regards,

Jobert T.

Intel Customer Support Technician