NVIDIA Reportedly Developing New Technology to Let SSDs Serve as Extra VRAM

According to media reports, NVIDIA is developing a new technology that could allow SSDs to act as additional VRAM in the future. When a graphics card runs out of VRAM, it could use data stored on a fast NVMe SSD instead, which could particularly benefit games with extremely high VRAM requirements or AI inference workloads. 

The feature is said to be based on NVIDIA's RTX IO technology and Microsoft's DirectStorage API.

Recently, NVIDIA officially announced the full open-sourcing of the cuFile API and its underlying storage software stack. The cuFile API is the core interface of NVIDIA's GPUDirect Storage technology, responsible for providing applications with the ability to directly access storage from the GPU. Its goal is to transfer storage data directly to the GPU within milliseconds, significantly shortening the data access path.

In the traditional process, data must first be temporarily cached in system memory before being copied to VRAM. This technology effectively reduces latency and significantly improves throughput efficiency through a "direct pass" approach. However, GPUDirect Storage is a technology that bypasses the CPU and system memory, enabling direct data exchange between storage devices and the GPU.

Working in tandem with this, RTX IO provides a direct data channel from NVMe SSDs to the GPU, allowing the GPU to bypass the CPU and system memory, read compressed data directly from the SSD, and decompress it on the GPU itself.

Although officially unconfirmed, the rumors reflect NVIDIA's exploration of software- and system-architecture-based solutions to address the growing tension between rising VRAM demands and hardware costs.