As of now, the PCI Express 4.0 standard has been finalized and officially released. The new protocol promises twice the per-lane bandwidth of PCI Express 3.0, allowing a GPU or other accelerator to transfer up to 64GB/s in a duplex x16 link. It’s also been a long time coming.
Initially, PCIe evolved relatively quickly. It debuted in 2003 with up to 8GB/s of bandwidth in full duplex mode (that’s 4GB/s of bandwidth in either direction, 8GB/s simultaneously). PCI Express 2.0 bumped that to 8GB/s unidirectional and 16GB/s bi-directional in 2005 and PCI Express 3.0 hit 32GB/s of bi-direction bandwidth (16GB/s in each direction) in 2010. PCI Express 4.0, however, had a much longer development period due to the difficulty of continuing to deliver a doubling of bandwidth in a backwards-and-forwards compatible electrical interface and mechanical form factor. Al Yanes, chairman and president of PCI-SIG, has released a blog post (PDF) detailing the advances of the new standard. New capabilities include:
- Extended tags and credits for service devices
- Reduced system latency
- Lane margining
- Superior RAS capabilities
- Scalability for added lanes and bandwidth
- Improved I/O virtualization and platform integration
In the past, we’ve often talked about PCIe strictly in terms of full-sized desktop GPUs, but that understates the benefit of these technologies. First of all, single GPU systems rarely get much benefit from higher PCI Express speeds, at least not initially. Multi-GPU systems, in contrast, can benefit more from these high-speed links because in these cases, PCIe is explicitly used for cross-GPU communication. We’ve put together a slideshow of the benefits and advances in PCI Express over the past few years, shown below. Each slide can be clicked to open it in a separate window.