life in the fast lane
While we haven’t seen PCIe 6.0 yet, the PCI SIG alliance has been busy drafting PCIe 7.0, and today they released our first look at PCIe 7.0’s capabilities. The new spec is unlikely to herald the arrival of ultra-fast GPUs, even if current generations of cards have plenty of bandwidth on PCIe 4.0 16x slots. This may pose interesting challenges for Phison and other SSD controller manufacturers who are busy saturating PCIe 5.0 connections, but the real driver of the new specification is networking.
The PCIe 6.0 specification describes a theoretical maximum bidirectional bandwidth of 256GB/s for a 16x slot, which is indeed impressive, but not enough to meet the needs of the growing LLM industry. The PCIe 6.0 16x slot can only support a single 800Gb/s network card, which is not enough to meet the needs of large machine learning clusters, leading Intel, NVIDIA, etc. to adopt their own solutions. The PCIe 7.0 16x slot will provide double the theoretical bandwidth and 512GB/s bidirectional bandwidth, which means we will break the TB barrier of 1.6Tb/s cards.That’s not quite enough to rival Nvidia’s NVLink or AMD’s Infinity Fabric interconnect, but it should keep PCIe viable for “slower” solutions
The increased bandwidth will also allow Compute Express Link to perform some interesting new tricks, which The Register notes will make PCIe-connected RAM more feasible. Take a look here to see how it works and its possible uses.