For decades, the benchmark for maximum PCIe performance was confined to the server chassis; namely, what kind of devices can be hosted directly by the server itself. If your accelerator was installed internally (into the designated add-in-slot/riser) the speed was, more or less, guaranteed. With the rise of AI and HPC composability, this paradigm has been broken . Historically, moving a GPU or high-speed storage outside  the server chassis resulted in a massive performance pe