192GB of VRAM in One PC… The Cheap Way

This video tests a dual-Intel GPU card to explore whether 192GB of VRAM in a single PC can be achieved affordably without relying on expensive NVIDIA hardware.

What it covers

Dual-Intel GPU Setup
The video examines the use of a dual-Intel GPU card to achieve high VRAM capacity in a single system.
192GB VRAM Feasibility
It evaluates whether having 192GB of VRAM in one PC is practical for real-world workloads like large LLM inference.
Cost vs. NVIDIA Alternatives
The build is compared to NVIDIA solutions, focusing on cost-effectiveness and performance trade-offs.
Performance for LLMs
The setup is tested for running large language models, assessing whether the VRAM capacity translates to real performance gains.
Hardware Limitations
The video discusses potential bottlenecks and limitations of using Intel GPUs for high-end AI workloads.
Build Challenges
It covers the technical challenges and compatibility issues encountered during the build and testing process.
Use Case Viability
The video concludes by assessing whether this approach makes sense for developers and AI researchers on a budget.