Recently, Credo Semiconductor, a listed company on Nasdaq in the United States that engages in AI computing chips, announced the completion of the acquisition of MicroLED optical interconnect innovation company Hyperlume, Inc. This case deserves our key interpretation, and this article will provide a detailed analysis.

1. CWho is redo? Why is high-speed Serdes the "traditional main force" of AI computing power?

1. Credo Semiconductor: "Hidden Champion" in the field of high-speed Serdes

Credo Semiconductor is a listed company headquartered in NASDAQ in the United States. It has been deeply involved in high-speed data interconnection chip technology for more than ten years and is a global leader in high-speed Serdes (serial/parallel converter) IP and chips. Its core business is to provide high-speed, low-power interconnection solutions for data centers, high-performance computing (HPC) and AI computing chips (such as GPUs and AI ASICs). Its customers include the world's leading cloud service providers, AI chip designers and optical module manufacturers. In the context of the explosive growth of AI computing power, Credo has become the "blood vessel" for data transmission between data centers and AI chips with its high-speed Serdes technology that has evolved from 200G/400G/800G to 1.6Tbps. When the GPU needs to exchange massive data with memory and other computing chips, Serdes is responsible for converting parallel signals into high-speed serial signals (or reverse conversion), and achieving cross-chip and cross-server interconnection through optical modules or copper cables..

2. High-speed Serdes: the "traditional main force" and "sweet burden" of AI computing power

As the number of AI model parameters exceeds 100 billion (for example, the number of GPT-4 parameters reaches 1.7 trillion), the demand for data I/O (input/output) rates of computing chips such as GPUs increases exponentially - a single GPU requires Tb/s (terabit per second) level of interconnection bandwidth to meet the real-time data exchange with memory and other GPUs during training or inference. Traditional solutions rely on high-speed SerDes channels: the rate of a single SerDes channel has reached 56Gbps, 112Gbps (PAM4 modulation format), but to "stuff" Tb-level bandwidth into the edge of the chip, hundreds or even thousands of SerDes macrocells must be lined up around the chip to form the so-called "Serdes Wall". This solution is extremely expensive:

• (1) Soaring power consumption: Serdes accounts for 10%-20% of the total chip energy consumption (for example, a high-end GPU has a total power consumption of 1000W, and Serdes may consume 100-200W);
• (2) Silicon area encroachment: Thousands of SerDes macrocells require huge chip edge space, squeezing the layout of computing units;
• (3) Packaging challenges: High-density Serdes wiring leads to a sharp increase in packaging difficulty and cost (such as PCB trace density, via loss, etc.).

In order to break through the bottleneck, the industry is trying two optimization paths:

seline;text-align:left;margin-left:0px !important;">
• SiGe Upconverter (RF upconversion): Modulates the tens of Gbps electrical signals output by Serdes to a radio frequency carrier (higher frequency) through SiGe (silicon germanium) technology, and uses frequency division multiplexing (FDM) to superimpose multiple frequency signals in a waveguide to increase the total bandwidth. However, this solution still relies on high-speed SerDes (the data needs to be serialized first).to tens of Gbps), the essence is to "strengthen the high-speed track" - Serdes is still the "bottleneck", but the transmission section is more efficient.
• MicroLED array (light domain parallel): As detailed below, this is the core related technology of Credo’s acquisition.

2. MicroLED optical communication: a "new track" that bypasses the "Serdes Wall"

1. What is MicroLED? Why can it subvert traditional optical interconnection?

MicroLED (microlight-emitting diode) is a micro-light-emitting device based on semiconductor materials (such as GaN, InGaN). The size of a single pixel can be as small as 1-10 microns (one-thousandth of traditional LED). In the field of optical interconnection, MicroLED directly converts electrical signals into optical signals through the principle of electroluminescence. Its core advantages are:

• (1) Ultra-low drive requirements: Each MicroLED only requires a few Gbps of medium and low-speed drive (traditional SerDes requires 56-112Gbps), and can be lit directly with a CMOS circuit without the need for complex ultra-high-speed SerDes circuits;
• (2) Spatial parallel expansion: By stacking channels in large-scale MicroLED arrays (such as hundreds to thousands of LEDs), the total bandwidth is increased in a "multi-water pipe parallel" manner (instead of relying on a single channel rate), breaking through the physical limitations of the "Serdes wall";
• (3) Ultra-small size and high integration: Micron-level light-emitting units can be directly integrated into chips or silicon interposers, greatly saving packaging space;
• (4) Low power consumption and long life: There is almost no energy loss in optical signal transmission (compared to the high Joule heat of electrical signals in copper wires), and the MicroLED luminescent material has high stability and a lifespan far longer than that of traditional electrical interconnect devices.

2. MicroLED vs. traditional solution: The essential difference between the two technical routes:

Properties	SiGe RF upconversion (traditional RF track)	MicroLED array (optical domain parallel track)
Whether to rely on high-speed SerDes	Dependent (data needs to be serialized to tens of Gbps before upconversion)	Not dependent (each LED only requires a few Gbps, driven directly in parallel)
Bandwidth expansion method	Frequency division multiplexing (increase carrier frequency points, such as f1/f2/f3 superposition)	Space division multiplexing (increase the number of LED channels, such as 1000 LEDs in parallel)
Single channel rate requirement	High (56-112Gbps level)	Low (1-5Gbps is sufficient)
Physical packaging bottleneck	Limited by the number of SerDes at the edge of the chip, power consumption and heat dissipation	No need for "Serdes wall", relying on array stacking to break through the I/O surface density
Applicable scenarios	Short-distance interconnection (<10cm inter-core RF interconnection)	Medium-distance/long-distance interconnection (through multi-core fiber array, such as GPU↔remote)

To put it simply, SiGe RF upconversion increases bandwidth through "frequency domain expansion", but essentially still relies on high-speed SerDes, but the transmission is more efficient; while MicroLED array expands bandwidth through "space domain parallelism", replacing a few high-speed SerDes with more low-speed LED channels, completely bypassing the "Serdes wall".

---

1. Acquisition background: Hyperlume's innovation and Credo's strategic needsHyperlume, Inc., which Credo acquired this time, is an innovative company focusing on MicroLED optical interconnection. The core of its technology is to achieve chip-level high-speed optical interconnection through large-scale MicroLED arrays, solving the bottlenecks of traditional optical interconnection in bandwidth, power consumption and integration. Hyperlume's solution complements Credo's original high-speed Serdes IP and CPO (co-packaged optics) technology - the former provides "optical domain parallel" light source and modulation capabilities, and the latter provides "electro-optical conversion" and "optical signal transmission" optimization experience.

2. Direct impact on AI computing power: from "electrical interconnection bottleneck" to "optical interconnection freedom"

• Break through physical limitations and increase interconnection bandwidth: through MDue to the "spatial parallelism" feature of the microLED array, computing chips such as GPUs no longer need to stack hundreds of high-speed SerDes. Instead, they can achieve Tb/s-level total bandwidth by integrating thousands of low-speed MicroLED channels (each channel only requires 1-5Gbps). For example, if 1,000 MicroLED channels (5Gbps each) are used, the total bandwidth can reach 5Tbps, and the chip edge area occupied is only 1/10 of the traditional Serdes solution.
• Reduce power consumption and cost: MicroLED's driving circuit is based on CMOS, which does not require the complex design of ultra-high-speed SerDes (such as high-swing drive, high-frequency balancing), and single-channel power consumption can be reduced by more than 50%; at the same time, the loss of optical signal transmission is much lower than that of electrical signals (copper wires), and the energy consumption of long-distance interconnections (such as the interconnection between GPUs in a data center) is further reduced.
• Promote the deep integration of "optical interconnection + AI": Credo's Serdes IP and Hyperlume's MicroLED array can be combined to build an "electrical-optical synergy" end-to-end interconnection solution - Serdes is responsible for low-speed data processing within the chip, and MicroLED is responsible for high-speed optical transmission between chips, ultimately achieving full-link high-speed interconnection from computing units to storage units to other GPUs. This solution is especially suitable for AI training scenarios (such as large model parameter exchange) and can significantly shorten training time.And reduce the overall energy consumption of the data center.

3. Industry impact: "Changing lanes and overtaking" opportunities in the AI computing power route

This acquisition marks the evolution of AI computing power interconnection technology from a single path of "depending on high-speed Serdes" to a diversified path of "electrical-optical hybrid parallelism". For the industry:

• Technical level: Accelerating the penetration of "optical interconnection" in AI chips, promoting the integration and innovation of Serdes and MicroLED technology (such as the integrated design of CPO co-packaging optics + MicroLED light source);
• Market level: With this acquisition, Credo has upgraded from a "high-speed Serdes supplier" to an "electro-optical full-stack interconnection solution provider", further consolidating its voice in AI computing infrastructure;
• Ecological level: It may drive related domestic industrial chains (such as MicroLED chip manufacturing, silicon photonic integration, optical module packaging), providing domestic AI chips with opportunities to "change lanes and overtake" (such as replacing imported high-speed Serdes solutions with domestic MicroLED light engines).

---

4. Summary: The future of AI computing power is hidden in the "synergy of light and electricity"

The case of Credo's acquisition of a MicroLED optical interconnect company is superficially a corporate merger, but in fact it reveals the deep logic of the development of AI computing power - when traditional electrical interconnection (Serdes) encounters physical limits, "optical interconnection" is becoming the key to breaking through technological innovation. The MicroLED array bypasses the "Serdes wall" through "spatial parallelism" and is combined with Credo's high-speed Serdes IP and CPO technology to build an efficient interconnection system of "electrical processing + optical transmission". It not only solves the bandwidth, power consumption and packaging problems of AI chips, but also lays the infrastructure foundation for more powerful computing power in the future (such as trillion-parameter large models, real-time AR/VR interaction). It can be predicted that as the technology matures and is commercialized, the integration of "optical interconnection + AI" will become the core of the next generation computing architecture, and the cooperation between Credo and Hyperlume may be the starting point of this change.. For the technology industry, this is not only an upgrade of technology, but also a profound reconstruction of computing efficiency, energy consumption and industrial structure.