The closer AI systems get to 224G per lane, the less forgiving copper starts to feel. At lower speeds, electrical links can absorb a surprising amount of ugliness in the channel before things really fall apart. Push into 224G PAM4 and that changes fast. Loss, reflections, rough copper, bad interfaces, all of it starts stacking up, and suddenly the old assumption that copper can just handle short-reach scale-up links begins looking less comfortable.
That is the gap MaxLinear is trying to hold open with Annapurna. Annapurna is a scale-up retimer used to maintain 224Gbps-per-lane PAM4 electrical connectivity in AI data center copper links. In a typical rack-scale AI system, this kind of device sits in the path between GPUs, switches, backplanes, or Active Electrical Cables, cleaning up the signal enough that copper can still be used where optics would otherwise start taking over.
What makes this interesting is not really the retimer itself. Retimers are already part of the conversation at these speeds. The real question is whether copper can keep its power, latency, and cost advantages inside scale-up fabrics once the channel impairments get severe enough that signal processing has to do much more of the heavy lifting.
224G Copper Stops Working On Good Intentions
MaxLinear is positioning Annapurna around a very specific problem. At 224G per lane, copper channels run into attenuation, return loss, inter-symbol interference, and other physical limits that make reliable reach harder to maintain. The PR also points to skin-effect resistance and copper roughness, which are the sort of details engineers usually end up fighting only after the link budget starts tightening more than expected.
That is where the DSP becomes central rather than helpful. MaxLinear says Annapurna uses adaptive equalization, transmit pre-emphasis, and proprietary DSP algorithms to compensate for those impairments and extend the usable electrical channel. That is really the story here. Copper is still attractive, but it is no longer coasting on material cost and familiarity alone. At 224G, it needs active help to remain viable.
The reason that matters in AI systems is fairly simple. Scale-up fabrics inside and between racks still care a lot about power per bit and latency, especially where tightly synchronized GPU communication is involved. Optical links remain the obvious choice for longer distances and scale-out networks, but they do not automatically win everywhere.
Scale-Up Fabrics Still Want Electrical Links
One of the more useful parts of this launch is how clearly it reflects the split inside modern AI infrastructure. Operators are not choosing between all-optical and all-electrical systems in some neat architectural vacuum. They are mixing both, using optics where reach dominates and copper where electrical links still make sense.
Annapurna is being aimed directly at that second category. MaxLinear says the device supports AEC and on-board retimer applications, with versions available in eight-lane and sixteen-lane configurations. That puts the parts into 1.6Tbps and 3.2Tbps design territory, which is exactly where interconnect choices start affecting tray layout, thermal budgets, and overall rack efficiency in ways that quickly stop being abstract.
There is also a protocol angle here. MaxLinear says Annapurna supports emerging scale-up environments including ESUN, UALink, and Ultra Ethernet. The point is not that one retimer solves the fabric. It is that the electrical link still needs to exist cleanly enough for those higher-level architectures to matter in the first place.
DSP Is Becoming Part Of The Copper Channel
This is probably the broader takeaway. AI infrastructure is not just pushing faster signaling into the same old interconnect assumptions. It is changing what counts as the channel. At 224G, the copper path is no longer just trace, connector, and cable. The signal conditioning silicon becomes part of the channel architecture as well.
That is what Annapurna represents. Not a last stand for copper, and not a replacement for optics either. More a reminder that electrical interconnects in AI racks are now being kept alive by increasingly serious DSP, because the economics and latency still make that worth doing in the parts of the system where reach has not yet won the argument.
Learn more and read the original announcement at www.maxlinear.com
Technology Overview
Annapurna is a scale-up retimer from MaxLinear for 224Gbps-per-lane PAM4 electrical connectivity in AI data center systems. It is designed for copper backplanes, Active Electrical Cables, and on-board retimer applications, using adaptive equalization, transmit pre-emphasis, and DSP to compensate for channel impairments. The device is offered in eight-lane and sixteen-lane versions for 1.6Tbps and 3.2Tbps applications.
Frequently Asked Questions
What is MaxLinear Annapurna used for?
It is used to extend 224Gbps-per-lane PAM4 copper connectivity in AI data center scale-up networks, including AEC and on-board retimer applications.
What bandwidth options does Annapurna support?
MaxLinear says Annapurna is available in eight-lane and sixteen-lane versions for 1.6Tbps and 3.2Tbps applications.
You may also like
