How PCIe Gen 5 SSDs and DirectStorage Are Rewriting Storage Architecture

Storage with throughput above 12 GB per second is often held back not by the media itself but by CPU decompression before the data ever reaches graphics memory.

The bottleneck moves

For years, storage was the slowest link in the chain. Fifth-generation NVMe flips that: the media can now deliver data far faster than the CPU can decompress it.

The traditional path passes through too many buffers. Data is read from the SSD into system memory, the CPU decompresses it, the result is copied back into system memory, then sent to VRAM. Every copy burns memory bandwidth, and decompression runs serially on just a few cores.

Moving the work to where the data is needed

DirectStorage shortens that chain by letting the GPU read compressed assets straight from the SSD and decompress them with its own thousands of parallel cores.

The fit is natural: decompressing many small blocks independently is a parallel problem, exactly the kind of workload GPUs are built for. CPU load drops sharply and wait times shrink to milliseconds.

Requirements that must be met together

The benefits do not appear just by installing a fast SSD. Software has to explicitly use the new interface, and assets must be stored in a compression format the GPU can decompress.

That is why many users who install a Gen 5 NVMe notice no difference at all: without support on the application side, the old path stays in use and the bottleneck does not move anywhere.

Beyond gaming

The same approach matters a great deal for data processing. Training a model on a dataset of hundreds of gigabytes means reading and decoding data continuously, and a slow data-loading pipeline leaves expensive accelerators idle, waiting their turn.

How well fast storage and compute accelerators work together is becoming the new measure of system performance — no longer the peak speed of each component in isolation.