Seeking Expert Views on AI-Accelerated Dual-GPU Rendering Architecture

In a previous post, I proposed a novel rendering architecture: using one GPU, such as an AMD RX 9070 XT, purely for fast 1440p or 4K rasterization, while a second dedicated AI accelerator (like an NVIDIA Tesla T4 or Ada L40) handles real-time upscaling to 8K using neural networks. The idea is to intelligently split workloads — raster on one card, upscaling on the other — rather than brute-forcing ultra-high resolutions through a single flagship GPU like the RTX 5090. Technically, this setup could leverage PCIe Gen 5 or NVLink to exchange frame buffers with minimal latency, and AI models like ESRGAN, Real-ESRGAN, or DLSS-like TensorRT pipelines could enhance frames in real-time, potentially within 3–6ms per frame. APIs like Vulkan or DirectX 12 could theoretically be used to direct the rendered output from the main GPU to system memory or shared VRAM, where the AI GPU takes over and performs intelligent super-resolution before sending the final frame to display. This method, if executed properly, could yield up to 300–400% performance gains by drastically reducing raster load, all while maintaining or even improving visual clarity. However, current consumer drivers and OS architectures do not natively support direct AI+render GPU collaboration, especially across vendors. Issues like frame latency, synchronization, and memory sharing (especially between AMD and NVIDIA) would require custom middleware or engine-level integration. Still, the core hardware already exists, and with high-speed interconnects and optimized inference runtimes (like ONNX or CUDA with pinned memory), this idea is more than feasible—it’s just waiting for the right software ecosystem.I’d really like to hear from professionals in graphics R&D, low-latency video AI, or game engine development—do you think a setup like this could realistically work as described? I’m curious about the practical possibilities and challenges involved. Feel free to share your insights at [e-mail]. — Muhammed Shafin P.