One of the pillars on which NVIDIA's dominance in the accelerator world rests is NVLink, a high-speed interconnect that allows chips to communicate directly not only within a single node, but also beyond its borders. AMD is trying to respond to this by promoting the XGMI/Infinity Fabric, and in its preliminary review of the Instinct MI300, issues of the topology of the red servers were raised.
Even then, at the time of the announcement of MI300, Broadcom announced support for this interconnect in future generations of its PCIe switches, and now the ServeTheHome resource has shared new details. XGMI will truly become switched, making it easier to scale systems based on AMD Instinct accelerators. The interconnect was officially named AFL (Accelerated Fabric Link).
AFL will continue to be based on PCI Express, in this case we are talking about PCI Express 7.0. Support for this technology will debut in Broadcom Atlas 4 PCIe switches. In addition to them, new Vantage 7 retimers will be released, which will also receive CXL 4.0 support.
But before that, Broadcom will begin shipping samples of Atlas 3 switch chips with 144 PCIe 6.0 lanes in the second half of 2024, and servers with such switches will appear in 2025. CXL support here will be extended to version 3.1.
As for retimers, Broadcom has already struck back at Astera Labs by announcing Vantage 5 and Vantage 6 series chips for the PCI Express 5.0 and PCI Express 6.0 ecosystems, respectively. They will be available in 8- and 16-lane variants with a bifurcation option and support for CXL 2.0 and 3.1.
Broadcom claims lower power consumption achieved through the use of a 5nm process technology, industry-leading SerDes blocks and advanced diagnostics integrated into the new retimers.
Cost-effectiveness plays an important role here: although even a 7-nm retimer consumes little, there are several such chips in each GPU server, which with further scaling results in very serious figures. In addition, less load will be placed on the cooling system, because while the CPU and accelerators can be serviced by a life support system, the remaining components in such servers are still cooled by conventional fans.
As for SerDes blocks, they will make it possible to lengthen connections by 40% while maintaining stable operation. Well, the availability of advanced diagnostic software with expanded capabilities will simplify the development, debugging and repair of new generation systems.
Vantage 5 retimers will be used with Atlas 2 switches in Broadcom solutions today, they will provide support for CXL 2.0, while Vantage 6 systems with CXL 3.1 support, as mentioned, should be released next year.
Astera Labs has something to worry about: if at the moment there is almost no alternative to its retimers, then in the near future the situation may radically change, since Broadcom has clearly realized the importance of this component in the PCI Express ecosystem and has estimated a solid size of the potential customer base.
If you notice an error, select it with the mouse and press CTRL+ENTER. | Can you write better? We always welcome new authors.
Source: