Ultra-High-Speed Optical Fiber Communication for Massively Parallel AI Processing

1Problem overview

In the last few years, large language models have grown rapidly in popularity, creating a large demand for high-performance computing among consumers and industry alike. As AI models have become significantly more advanced, they have also become increasingly complex. Today, AI training workloads can require thousands of flagship, high-power GPUs. These GPUs are configured in servers to maximize bandwidth and create fast connections from GPU to GPU, so they can constantly exchange model parameters, gradients and training data. Because of this data transfer requirement, the system connecting the GPUs has become nearly as important as the GPUs themselves.

This GPU-to-GPU communication has become much more complicated in recent years. Today, communication occurs across multiple “layers” within an AI computing architecture, from communication within a rack to rack-to-rack, pod-to-pod and finally building-to-building. Technologies such as NVLink Gen 5 have heavily optimized communication within servers and racks; NVIDIA describes one NVL72 rack as “essentially one giant GPU.”¹ Communication between pods and buildings, however, still has significant room for improvement. It is commonly handled with Ethernet or InfiniBand, in large deployments built on fat-tree and Clos network architectures. These provide high performance, but often at the cost of increased complexity and reduced scalability as systems continue to grow.

As AI systems grow, more data must be moved from rack to rack, pod to pod and building to building. As physical link lengths increase, latency, BER and power consumption can increase significantly, and each additional network layer built on Ethernet or InfiniBand adds complexity while reducing overall scalability. Consequently, many companies are searching for faster, more efficient and more scalable methods of data communication.

2Why optical fiber

One of the most promising solutions is optical fiber communication. Optical fiber can provide significantly higher bandwidth than traditional copper connections while maintaining very low attenuation over long distances. A technology known as wavelength-division multiplexing (WDM) allows multiple independent data channels to operate simultaneously on a single fiber, which improves scalability while helping reduce overall power consumption.

Because future AI systems will require scalable, high-bandwidth communication between racks, pods and buildings, optical networking architectures have emerged as a promising way past the limits of traditional electrical interconnects.

Key terms

Single-mode fiber

Fiber that carries a single light path. It loses about 0.2 dB per kilometer near 1550 nm, while passive copper reaches only a few meters at 100+ Gb/s per lane.

Dense WDM (DWDM)

WDM with channels spaced 100 GHz (about 0.8 nm) or 50 GHz apart in the C-band (1530-1565 nm), giving 40 to 80 or more channels on one fiber.

PAM4

A signal with four levels, carrying 2 bits per symbol, so 53 GBd gives about 106 Gb/s per lane. Each eye opening is a third as tall as in two-level signaling, a penalty of about 9.5 dB in SNR,¹⁴ which is why these links use forward error correction.

Coherent optics

Transmission that encodes data in the amplitude, phase and polarization of light, with a DSP chip correcting fiber distortion. It carries 800 Gb/s to 1.2 Tb/s per wavelength.⁵

Optical circuit switching (OCS)

A switch that steers light from any input fiber to any output fiber without converting it to electrical (O-E-O). Reconfiguring takes milliseconds, which suits long-lived traffic such as AI training. Google has used OCS in production since TPU v4.⁸

3Existing solutions

Many companies have invested significant time and resources into researching AI communication architectures. NVIDIA has developed systems using NVLink and NVSwitch for high-bandwidth communication within servers and racks, and commonly uses InfiniBand in large deployments built around fat-tree (Clos) architectures. Google has approached the challenge through TPU pods, which use torus-based interconnects designed to support distributed AI training. Microsoft has addressed similar challenges with Ethernet leaf-spine architectures. Each of these approaches supports modern AI workloads; however, as AI systems continue to grow in size and complexity, these solutions become increasingly expensive and difficult to scale (Table 1).

Table 1.Existing approaches to GPU networking at scale.

Optical networking technologies are already widely used throughout telecommunications and internet backbone infrastructure, and technologies such as WDM, coherent optics and photonic integrated circuits are being actively researched for future AI networking. These trends point to a growing demand for scalable optical communication systems that can support next-generation AI workloads, which motivated the architecture in this project. Recent industry developments:

• 2023. Google publishes TPU v4, the first production supercomputer with reconfigurable optical circuit switching.⁷
• Feb 2025. Nokia completes its acquisition of Infinera, the maker of this project’s Layer 4 transport hardware.¹⁵
• Mar 2025. NVIDIA announces Spectrum-X and Quantum-X Photonics switches, with the optics co-packaged inside the switch package, claiming 3.5× better power efficiency.⁹
• 2025. HUBER+SUHNER opens a new POLATIS optical-circuit-switch factory for AI data centers and plans at least 5× more capacity.¹⁶

4The architecture

Our objective was to develop a scalable communication architecture for large AI computing systems. The design combines pre-existing commercial products and established networking concepts into one unified, scalable model with four layers (Figs. 1 and 2). Layer 1 uses NVIDIA’s GB300 compute trays as the primary compute layer, and Layer 2 uses NVIDIA’s NVL72 rack architecture for high-bandwidth copper connections between those trays. Both are preconfigured and deeply optimized by NVIDIA already. Layers 3 and 4 are where our research was primarily focused: a HUBER+SUHNER POLATIS Series 7000 optical circuit switch connects racks and pods inside the building, and the Nokia (formerly Infinera) 1830 GX G42 with FlexILS DWDM transport carries data between buildings.

To evaluate the design, we considered bandwidth, latency, SNR and BER, along with cost, reliability, scalability and implementation complexity.

Table 2.Hardware and estimates for each layer.

5Following the data

L1 Compute. Data starts processing at Layer 1, the GB300 compute tray. Each tray has 4 B300 Blackwell GPUs wired to 2 Arm-based Grace CPUs, which communicate over NVIDIA’s NVLink-C2C, an ultra-high-bandwidth, low-latency chip-to-chip connection. These trays are vital to the entire architecture: the rest of it is designed to connect them to each other on a massive scale. Because of that scale, the bottlenecks sit in Layers 2 to 4, which are designed to reduce them as much as possible.

L2 Rack. At Layer 2, each compute tray connects to the NVL72 rack spine. The spine uses NVIDIA’s NVLink Gen 5 to likewise maximize bandwidth and minimize latency, clustering 18 trays (72 GPUs) into one rack. This layer is already designed and optimized by NVIDIA. Data leaves the NVL72 rack as 800 Gb/s PAM4 over single-mode fiber at 1310 nm.

L3 Pod optical fabric. At Layer 3, we selected the POLATIS Series 7000 optical circuit switch, specifically to maximize bandwidth. It receives all the aggregated traffic from the NVL72 racks, connecting racks within a pod and pods within the building. Because the switch is entirely optical, the signal is never converted to electrical, which keeps signal quality high. The switch is also non-blocking, allowing multiple optical paths to exist at once without creating network congestion. This capability is primarily why we chose this model. Layer 3 outputs the same single-fiber PAM4 signal at 1310 nm that came in, adding about 25 ns.⁴

L4 Optical transport. Finally, at Layer 4 the objective switches from low-latency switching to long-distance data transmission. The PAM4 data goes into the Nokia 1830 GX G42 transport platform, where ICE-X coherent optics convert it into DWDM channels for transmission to a remote building about 10 km away, and FlexILS combines those channels onto one fiber. This is the only optical-electrical-optical conversion in the design, and it applies only to traffic leaving the building.

Light travels through glass at about 2×10⁸ m/s (5 ns per meter), so over building-scale distances, propagation is the largest part of the latency.

6Engineering analysis

Jonas Craig led the engineering calculations. Each layer was analyzed with first-order hand calculations of bandwidth, SNR, BER, optical loss and latency, using standard textbook models, to check that the design is feasible.

Bandwidth came out to 112 Gb/s per lane for Layer 1 and 800 Gb/s for Layers 2 and 3. These are per-link (per-GPU) figures.

Table 3.Per-layer estimates.

Limitations

• Noise is modeled as thermal (plus shot noise at Layer 2). Real optical links are also limited by laser intensity noise, receiver amplifier noise and, over distance, optical amplifier noise, all of which lower SNR further.
• Real 800G links run at a raw (pre-error-correction) BER far above 10⁻¹² and rely on forward error correction (FEC) to reach the ~10⁻¹² target.
• The OCS loss uses 1 dB; the POLATIS datasheet lists 1.5 dB typical and 2.7 dB maximum.⁴ Fiber loss at 1310 nm is closer to 0.35 dB/km than 0.2. Over ≤1 km, both change the Layer 3 budget by less than 2 dB.

7Scale and cost

18 trays per rack, 8 racks per pod and 8 pods per building: 64 racks, 1,152 compute trays and 4,608 GPUs.

Table 4.Estimated hardware cost.

At roughly $180-250 million, a system at this scale is marketable only to very large organizations, such as hyperscale cloud providers and national laboratories, rather than the average data center company. It could be scaled down by condensing the layers and using less hardware, but with a significant loss in performance for the end user.

8Energy and impact

Nicholas Gazzara led the environmental impact and infrastructure analysis. The architecture has both sustainability benefits and broader societal impacts. By using optical networking technologies like optical circuit switching (OCS) and dense wavelength-division multiplexing (DWDM), the system can transmit large amounts of data more efficiently, reducing the need for additional physical infrastructure and lowering network power consumption. However, the system still operates at a very large scale. Each NVL72 rack draws roughly 120 kW,¹⁷ so 64 racks and 1,152 compute trays come to about 8 MW of continuous computing load and significant heat. That requires advanced cooling and increases both energy and water use, making efficiency a critical factor in the design.

As data centers continue to expand to support AI workloads, their impact on electricity use and carbon emissions grows. U.S. data centers used 58 TWh in 2014 and 176 TWh in 2023, about 4.4% of the national total.¹⁰ Federal projections put national data center use at 325-580 TWh by 2028, between 6.7% and 12% of U.S. electricity (Fig. 5).¹¹ While systems like this enable major advances in research, healthcare and automation, they also show the importance of balancing high-performance computing with responsible energy use and long-term sustainability.

9Design decisions and next steps

1. Build on NVIDIA’s rack. Layers 1 and 2 are preconfigured and deeply optimized by NVIDIA already, so our research focused on Layers 3 and 4, where communication between pods and buildings still has room for improvement.
2. Optical circuit switching at Layer 3. We chose the POLATIS switch specifically to maximize bandwidth. It is entirely optical, so signal quality is preserved, and it is non-blocking, so many paths can be open at once. Reconfiguring takes about 50 ms,⁴ so the switch suits long-lived flows; the full system keeps an InfiniBand cluster network for traffic that needs per-packet switching.
3. Coherent DWDM at Layer 4. Although NVIDIA hardware was initially considered, we chose Infinera (now Nokia) optical transport for its scalability. At building scale, the objective switches from low-latency switching to long-distance transmission, and DWDM puts many 800 Gb/s channels on a single fiber.
4. Commercial parts only. The design combines pre-existing commercial products and established networking concepts into one unified, scalable model that can be built with hardware available today.
5. Next steps. A full-size version needs to be tested with all the servers in place, to measure how much electricity the system draws while running AI workloads and how much heat it adds. The analysis should also be extended to a full link budget that includes all major noise sources and forward error correction, along with a power model for each layer.

10Contributions

Zach Busch led the system design, defined the four-layer hierarchy, selected the hardware for each layer, worked out how data moves between the layers, and built the architecture diagram.

Jonas Craig led the engineering calculations of bandwidth, SNR, BER, optical loss and latency for each layer.

Nicholas Gazzara led the environmental impact and infrastructure analysis, covering cost, power, cooling and energy use.

Course. ENGR 113 First-Year Engineering Design, Drexel University

Instructor. Dr. Afshin S. Daryoush

Result. Placed 2nd of 10 teams, ranked by the instructor.

References

1. NVIDIA, “GB300 NVL72.” nvidia.com (opens in a new tab)
2. NVIDIA, “NVL72 AI Factory reference architecture: system hardware and components.” docs.nvidia.com (opens in a new tab)
3. HUBER+SUHNER, “POLATIS 7000 Optical Circuit Switch.” hubersuhner.com (opens in a new tab)
4. HUBER+SUHNER, “POLATIS Series 7000n datasheet.” lasercomponents.com (opens in a new tab)
5. Nokia, “1830 Global Express (GX).” nokia.com (opens in a new tab)
6. Nokia, “1830 Flexible Intelligent Line System (FlexILS).” nokia.com (opens in a new tab)
7. N. P. Jouppi et al., “TPU v4: An Optically Reconfigurable Supercomputer for Machine Learning with Hardware Support for Embeddings,” ISCA 2023. arxiv.org (opens in a new tab)
8. Google Cloud, “TPU v4 enables performance, energy and CO2e efficiency gains.” cloud.google.com (opens in a new tab)
9. NVIDIA, “NVIDIA Announces Spectrum-X Photonics, Co-Packaged Optics Networking Switches…,” Mar. 18, 2025. investor.nvidia.com (opens in a new tab)
10. A. Shehabi et al., “2024 United States Data Center Energy Usage Report,” LBNL, Dec. 2024. eta-publications.lbl.gov (opens in a new tab)
11. U.S. DOE, “DOE Releases New Report Evaluating Increase in Electricity Demand from Data Centers.” energy.gov (opens in a new tab)
12. D. Foley and J. Danskin, “Ultra-Performance Pascal GPU and NVLink Interconnect,” IEEE Micro, 2017. doi.org (opens in a new tab)
13. C. Gómez et al., “Beyond Fat-tree: Unidirectional Load-Balanced Multistage Interconnection Network,” IEEE CAL, 2008. doi.org (opens in a new tab)
14. Keysight, “PAM4 Design Challenges and the Implications on Test.” keysight.com (opens in a new tab)
15. Converge Digest, “Nokia completes Infinera acquisition,” Mar. 2025. convergedigest.com (opens in a new tab)
16. Signal Integrity Journal, “HUBER+SUHNER opens new POLATIS optical circuit switch production site to support AI data centers,” 2025. signalintegrityjournal.com (opens in a new tab)
17. Introl, “Why the NVIDIA GB300 NVL72 (Blackwell Ultra) Matters.” introl.com (opens in a new tab)