It looks as though
AMD will be introducing a new GPU-to-GPU interconnected when its
Vega 20 GPUs arrive on the scene, presumably by the end of the year. The new interconnected, called XGMI, is based on the company's Infinity Fabric and is AMD's answer to
NVIDIA's NVLink, the latter of which is part of the
GeForce RTX series.
"We have already taped out multiple 7nm products at TSMC, including our first 7nm GPU planned to launch later this year and our first 7nm server CPU that we plan to launch in 2019. Our work with TSMC on their 7nm node has gone very well and we have seen excellent results from early silicon," AMD's Mark Papermaster stated in a blog post.
At this point, we still have limited information about Vega 20. There have been a few leaks pertaining to
workstation cards based on Vega 20, and is likely where the XGMI interconnect will show up. How do we know its coming?
The folks at Phoronix dissected a new set of Linux patches and found references to the interconnect. There are not details to glean, other than Vega 20 apparently supporting XGMI. However, previously leaked slides have indicated that both XGMI and PCI Express 4.0 would be supported on Vega 20, and the new Linux patches appear to confirm this.. It's also expected that AMD's Naples server CPU architecture will support XGMI when it arrives.
It makes sense for AMD to dish out XGMI on the professional/server side before rolling it out to consumers.
NVIDIA took the same approach with NVLink—it first appeared in its Tesla HPC products before reaching consumers.
The performance comparison between XGMI and NVLink will be interesting. As it stands, NVIDIA's
Tesla V100 offers up to 300GB/s of interconnect bandwidth via NVLink, which is 10X the bandwidth of PCIe Gen 3.