Technical Article Analysis | Broadcom 200G VCSEL Deep Dive: NPO, the Power-Saving Weapon for AI Scale-Up Networks

While much of the industry is still talking about 100G per lane, Broadcom has already planted its flag on 200G VCSELs. As AI compute demand explodes, copper (DAC) reach at 200G shrinks to a meager 1 to 2 meters, a nightmare for rack designers. In this issue we take apart this major Broadcom paper to see how it uses the low power and high reliability of VCSELs to build a lifeline for AI scale-up networks: Near-Package Optics (NPO).

• Title: 200G VCSEL Development and Proposal of Using VCSELs for Near-Package-Optics Scale-Up Application

  
  

• Authors: Tzu Hao Chow, Jingyi Wang, Sizhu Jiang, M. V. Ramana Murty, et al.

  
  

• Affiliation: Broadcom Inc.

  
  

• Published in: Photonics 2026 (MDPI)

  
  

Figure-by-figure analysis

The paper contains 12 figures and 1 table, building a complete chain of logic from the single device up to the system application:

• Figure 1. VCSEL technology evolution

• Details: Shows the exponential growth in VCSEL modulation rates from 1 Gb/s in 1995 to 200 Gb/s today.

  
  

• Why it matters: It marks 200G as the current technology frontier, with Broadcom using PAM4 to push single-lane bandwidth to new highs.

  
  
  
  
  
  

• Figure 2. 200G FEC code word error rate

  

  • Details: A 9-hour traffic test over 50 meters of OM4 fiber showed no uncorrectable errors.

    
    

  • Why it matters: It demonstrates the high stability of 200G VCSELs over long reach (by multimode fiber standards).

    
    

• Figure 3. L-I and V-I characteristics

  

  • Details: At 9 mA bias, optical power exceeds 3 mW with a drive voltage of about 2.5 V.

    
    

  • Why it matters: This is the device's "health check," showing the design keeps optical output efficient at low current, which helps reduce heat dissipation.

    
    

• Figure 4. S21 frequency response

  

  • Details: At 9 mA, the -3 dB bandwidth exceeds 35 GHz, with a flat response and good damping.

    
    

  • Why it matters: Bandwidth is the foundation of 200G modulation. Broadcom squeezes out top-end high-frequency performance by reducing parasitic inductance and capacitance.

    
    

• Figure 5. Relative intensity noise (RIN)

  

  • Details: Measured RIN is below -152 dB/Hz.

    
    

  • Why it matters: PAM4 signals are very sensitive to noise. Low RIN means the laser's "breathing" is quiet, keeping the signal clean.

    
    

• Figure 6. 200 Gbps PAM4 eye diagrams

  

  • Details: Compares eye diagrams over 2m, 30m and 50m of fiber, with the 50m case using custom high-bandwidth fiber from Corning.

    
    

  • Why it matters: The eyes are clearly open, showing modulation quality good enough for 200G operation across 50 meters.

    
    

• Figure 7. Normalized output power vs. service life

  

  • Details: At 60°C, predicted power degradation after 25 years of use stays well below the 20% failure threshold.

    
    

  • Why it matters: It addresses market concerns that high-speed VCSELs are short-lived, and shows Broadcom's deep expertise in reliability.

    
    

• Figure 8. Network equipment disaggregation

  

  • Details: Shows XPU servers and switch racks separated by optical links.

    
    

  • Why it matters: Longer optical reach gives data centers more flexibility in placing cooling and power resources.

    
    

• Figure 9. 3.2T VCSEL NPO engine concept

  

  • Details: Integrates four 8-channel VCSEL and PD arrays in a compact 19mm x 19mm footprint.

    
    

  • Why it matters: This is the paper's core proposal: moving the optical transceiver from a pluggable module to right next to the ASIC, sharply shortening the electrical path.

    
    

• Figure 10. S21 response of a 100G VCSEL at high temperature

  

  • Details: Shows bandwidth holding above 32 GHz even at 75°C.

    
    

  • Why it matters: It mimics the harsh thermal environment inside a server, making sure the system does not "go weak at the knees" under heat load.

    
    

• Figure 11. Layout of 18 NPO engines

  

  • Details: Delivers 73.7 Tbps of total bandwidth within a 120mm x 60mm area.

    
    

  • Why it matters: It demonstrates very high escape bandwidth density (beachfront bandwidth density) of 0.6 Tbps/mm.

    
    

• Figure 12. XPU-to-NPO insertion loss model

  

  • Details: Breaks down signal loss in detail from the chip bump to the PCB trace.

    
    

  • Why it matters: It explains NPO's advantage: loss is held to 2-7 dB, far better than the 8-16 dB of conventional pluggable modules.

    

Conclusion

This paper reveals the backbone of future AI networks: far from obsolete, VCSELs are shining in the 200G era. Compared with roughly 5 to 10 pJ/bit for silicon photonics, Broadcom's VCSEL NPO approach needs only about 1 pJ/bit.

By sitting right up close to the ASIC, the VCSEL NPO solution eliminates the power-hungry DSP chip, which not only cuts cost dramatically but also brings power down to an enviable level. For AI scale-up networks that interconnect thousands upon thousands of GPUs, this 1 pJ/bit efficiency could be what decides whether a data center's power bill blows up.