GPUs and AI

NVMe

Also called: NVMe SSD, M.2 NVMe

A protocol that connects an SSD directly to the PCIe bus. An NVMe SSD reads and writes much faster than a SATA SSD, which speeds up data loading.

NVMe, short for Non-Volatile Memory Express, is the communication protocol of modern SSDs. Instead of going through the SATA interface inherited from hard drives, the SSD is connected directly to PCIe lanes. The most common form factor is the small M.2 stick that screws onto the motherboard; servers also use U.2 or E1.S form factors, which are hot-swappable.

SATA tops out at around 550 MB/s. A PCIe 4.0 NVMe SSD exceeds 7 GB/s, and PCIe 5.0 models go beyond 10 GB/s. NVMe also handles a very large number of command queues in parallel, which sharply reduces waiting times when many files are read at once.

For AI, NVMe storage matters mainly at load time: a language model weighs anywhere from a few gigabytes to several hundred, and it has to be transferred from the drive to the GPU's memory every time it starts. For training, it feeds the GPUs with data without keeping them waiting. On a product page, look at the capacity, because models pile up quickly, the PCIe generation and, for a server, the endurance expressed in terabytes written (TBW). For a mining machine, on the other hand, modest storage is enough.

The Threadripper AI workstation we offer comes with 4 TB of NVMe storage, enough to keep plenty of models.