Google's 448Gbps Roadmap: AI Training Forces PAM6/PAM8 Over PAM4 for Electrical Lanes
Google's OIF workshop report analyzes how AI model training across tens of thousands of xPUs drives 448Gbps/lane interconnects, comparing NVIDIA NVLink and Google ICI architectures, evaluating PAM4 through PAM12 modulation trade-offs, and proposing CPC co-packaged copper and OSFP3.2T modules as key hardware solutions for scale-up and scale-out networks.
AI Clusters Drive 448Gbps High-Speed I/O
At the OIF 448Gbps AI Workshop, Google presented The Path to 448Gbps , rooted in Google Cloud AI cluster practice. The report analyzes large-model training demand for high-speed I/O and compares NVIDIA GPU and Google TPU scale-up/scale-out interconnect architectures. The fundamental driver is not device-technology iteration but AI training workloads pushing I/O upgrades: tens of thousands of xPUs training together make chip-to-chip interconnect bandwidth the cluster performance bottleneck, creating the need for 448Gbps/lane signaling.
New-generation AI chips — Google Ironwood (TPU v7), NVIDIA Grace Blackwell Ultra, Vera Rubin — require tens of thousands of xPUs per cluster. Clusters split into two networks:
Scale-Up (tightly coupled super-node): intra-subset xPU ultra-high-bandwidth interconnect.
Scale-Out (horizontal expansion): cluster-wide fabric.
Both network interfaces land on xPU package pins, placing enormous pressure on chip I/O.
Two Mainstream Cluster Architectures
1. NVIDIA GPU Cluster
Inside the rack, NVLink handles Scale-Up with NVSwitch. NVL72/NVL576 racks use copper interconnects. Copper wins on power and cost but faces physical limits on interconnect radius and GPU count.
2. Google TPU Cluster
Uses ICI (Inter-Chip Interconnect) with no external switches. The base unit is an elemental cube of 64 chips; intra-rack 3D torus runs over copper. Ironwood (TPU v7) SuperPod scales to 9,216 chips; inter-unit links use optical modules plus OCS (optical circuit switches).
448Gbps Candidate Modulation Evaluation
Google compared PAM4, PAM6, PAM8, and PAM12 at 448Gbps, examining Nyquist frequency and SNR required for different symbol error rates:
PAM4: 106 GHz Nyquist — excessive bandwidth pressure on chip and module electrical channels; viable for optical fiber channels.
PAM6: 85 GHz Nyquist — bandwidth pressure reduced, SNR penalty increased; a compromise candidate.
PAM8: 70 GHz Nyquist — SNR requirement rises further.
PAM12: 60 GHz Nyquist — SNR conditions too stringent; crosstalk and channel impairment make engineering implementation extremely difficult.
Conclusion: For chip-to-module electrical channels, PAM4 bandwidth pressure is too high and PAM12 SNR threshold is unrealistic; PAM6 and PAM8 are the primary evaluation targets. On the optical fiber side, PAM4 remains usable.
System Interconnect Hardware Options
1. Flyover Cables
Route signals over the PCB to avoid trace loss, improving channel quality, though engineering implementation remains challenging.
2. CPC (Co-Packaged Copper)
Signals exit from the package top, bypassing BGA balls, vias, and PCB — significantly improving channel quality.
I/O density doubles using 0.5 mm pitch elastic connectors versus conventional 1.0 mm BGA.
Switch ASICs can also adopt CPC, enabling 128×3.2T port switches at 448Gbps.
3. OSFP3.2T and MSA Multi-Source Module Concepts
Target 8×448Gbps; DAC, LPO, and TRO link budgets fall short.
In-rack priority goes to AEC (active electrical cables) and AOC (active optical cables); IM-DD direct-detect optical reach is limited (<1 km), requiring Coh-lite (lightweight coherent) support.
Module power supply upgrades to 12V, supporting up to 50W.
Scale-Up/Scale-Out networks need not be backward compatible with legacy form factors; optimize for signal margin and power delivery first.
AI workload demand drives 448Gbps links. xPU trays and switch trays favor front-panel pluggable form factors to balance deployment density, reduce operational risk, and increase architectural flexibility. Chip-to-module electrical channels still require continuous technical improvement to enable 448Gbps deployment.
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
