Ahead-looking: The AI world wants lots of knowledge, and that knowledge wants to maneuver a lot sooner onto next-gen storage {hardware}. The PCI Particular Curiosity Group is engaged on PCIe 6.0 and seven.0 requirements to ease the ache with knowledge throughput, although a number of points nonetheless should be ironed out earlier than normal deployment.
Final yr, Micron teased its first PCIe 6 SSD, promising spectacular bandwidth charges. Extra lately, the US reminiscence producer has partnered with swap maker Astera Labs to indicate the brand new drives in motion. The 2 firms revealed what a correct PCIe 6 structure can supply throughout a DesignCon 2025 showcase. It was the primary public demonstration of end-to-end interoperability between a PCIe 6.x swap and a PCIe 6.x SSD.
The DesignCon demo featured a Micron “Information Heart” SSD linked to a Scorpio P-Sequence Material Change developed by Astera. Collectively, the 2 new applied sciences surpassed a sequential learn velocity of 27GB/s. In 2024, Micron claimed its PCIe 6 drive would supply as much as 26 GB/s sequential learn speeds, so it exceeded that brag.
Astera notes that the Scorpio P-Sequence Material Change is the primary designed particularly for PCIe 6 units. The custom-built swap can work with 64 PCIe 6.x lanes and a 4-port structure. It’s simply as quick in managing GPU, CPU, SSD, and NIC knowledge flows. The PCIe 6 and PCIe 7 specs aren’t fairly prepared for prime time, as firms are nonetheless in search of methods to keep away from velocity throttling as a result of extreme warmth generated by the brand new bus tech.
Reaching over 27 gigabits per second required a artistic connection administration method. Astera linked its swap to an unspecified CPU, an Nvidia H100 GPU, and two Micron PCIe 6.x E3.S SSDs. The demo used Nvidia’s Magnum IO GPUDirect Storage expertise to ascertain a direct knowledge path between the GPU and the SSDs.
Astera Labs claims that PCIe 6.x is changing into required to construct higher AI and cloud infrastructures. The brand new customary can double the bandwidth of PCIe 5.0 units, delivering as much as 256 GB/s of bidirectional knowledge charges per x16 lane.
The low-latency, high-speed structure ought to feed sufficient knowledge to satiate AI-related workloads. The demo seemingly proves that PCIe 6.x is now prepared for knowledge middle deployment. Nevertheless, conventional PC merchandise mustn’t anticipate PCIe 6-compliant motherboards anytime quickly.
