Marvell XPUs and Nvidia (NVLink Fusion)
Marvell custom XPUs are built to work with Nvidia hardware, not only as a standalone replacement.
Nvidia opened NVLink Fusion so third-party accelerators can plug into its high-bandwidth NVLink fabric. Marvell is an official partner. In that setup:
• A Marvell XPU can sit in the same rack-scale system as Nvidia GPUs
• They talk over NVLink (very high bandwidth, low latency)
• Nvidia still supplies pieces such as Vera CPUs, ConnectX NICs, BlueField DPUs, NVLink switches, and Spectrum-X Ethernet
• The cluster is required to include at least one Nvidia component
Nvidia also put $2 billion into Marvell and they are collaborating on silicon photonics / optical interconnect. So Marvell XPUs are explicitly meant to mix with Nvidia products in hybrid “AI factory” designs (training + inference). That is the official path for customers who want custom silicon and Nvidia’s interconnect and software stack.
Broadcom
Broadcom also designs custom AI chips (XPUs / ASICs) for big customers — Google-style TPUs historically, plus more recent work with OpenAI (Jalapeño inference chip), Meta, and others. Those chips are not in the NVLink Fusion program the way Marvell’s are.
@__Con_ Yes its what I think the true dip is within first ten days oct I just think last 2 weeks of sept we see some red as well but not the dip or main red