Add an optional Phison aiDAPTIV+ card to expand VRAM memory of FutureProofHomes's AI Base Station?

[Hedda]

Not sure of the price of those Phison aiDAPTIV+ adapter but wondering it could make it optional to use them (or rather their middleware software) in your upcoming AI Base Station product:

• https://www.cnx-software.com/2025/01/27/phison-aidaptiv-ai-solution-uses-ssds-to-expand-gpu-memory-for-large-language-model-training/

As I I understand are really more or less just a fast SSD/NVMe disk that comes with the advanced "aiDAPTIVLink" middleware library (driver software which they call "aiDAPTIVCache") that actually does all the work of enabling that SSD/NVMe to act kind of like a virtualization hypervisor for the GPU that will allow it to function as a VRAM swap space, i.e. VRAM cache/buffer for doing memory paging of VRAM), and essentially exposing it all to the application as "virtual VRAM".

"This improvement is due to Phison’s middleware, which optimizes GPU memory by slicing the model, holding pending slices in aiDAPTIVCache, and swapping slices efficiently between the cache and the GPU."

At least I assume that the Phison aiDAPTIV adapters themselves are actually just ordinary SSD/NVMe disks and you could probably use any other disk instead but they simply choosen to package the software license with their own hardware as a kind of hardware security key.

The reason for this idea is that believe that your upcoming AI Base Station will probably be based on Nvidia Jetson Orin NX 16GB if is released soon, and if that is the case then its 16GB of VRAM would infortunatly make it rather limited, and even if using NVIDIA Jetson AGX Orin 32GB it would still be limited, so having a lot more "virtual VRAM" would be very nice! 😉

That linked article mention this; "ADLINK’s newly launched DLAP Supreme series uses Phison’s aiDAPTIV+ technology to solve memory and performance limitations in edge devices for generative AI. The DLAP Supreme series achieves 8x faster inference speeds, and 4x longer token lengths, and supports large language model training, even on devices with limited memory configurations like NVIDIA Jetson AGX Orin 32GB using the Gemma 27B model."