In this exclusive interview, GlobalFoundries representative Vikas Gupta discussed the advantages of silicon photonics for developing efficient AI chips.

Multinational semiconductor manufacturer GlobalFoundries (GF) recently introduced new silicon photonics chips, ICs that integrate optical elements on a silicon substrate. These chips could help address the exponential data requirements of AI tools.

In this exclusive All About Circuits Q&A, we spoke with Vikas Gupta, senior director of product management at GF, about the new chips and how silicon photonics could help to meet the rising demands of AI.

Vikas Gupta, senior director of product management at GF. 
 

What is your professional background? How did it lead you to your role at GF? 

Vikas Gupta: I am the senior director of product management for the silicon photonics business line at GF. My team manages silicon photonics technology and packaging and also interacts with some of the lead adopters of silicon photonics (SiPh).

This is my second stint at GF. Before re-joining the company, I was the vice president of design systems at POET Technologies and worked on optical interposers and VCSELs. Prior to POET Technologies, I managed the GF PDK and modeling teams worldwide. I also held various engineering and management roles at AMD, Xilinx, and Texas Instruments.

How has GF helped to develop cutting-edge AI chips in recent years?

Gupta: Typical AI/ML hardware systems include large clusters of graphics processing units (GPU) and memory with networking interfaces to interconnect the GPUs and the rack switch. Photonic integrated circuits (PICs) can make these systems a lot more efficient by:

• Replacing copper interconnects with photonic interconnects
• Disaggregating the memory for efficient use
• Replacing the electronic compute with photonic compute chips

Processing power has increased by 60,000x over the last decade. During the same period, the dynamic random access memory (DRAM) bandwidth increased by 100x, and the network bandwidth increased only by 30x.

Data centers need better resources to efficiently compute the massive amounts of data associated with artificial intelligence (AI) systems.

One can conclude that the immediate value of PICs is to address the networking bandwidth challenges, overcoming the limitations of copper-based interconnects. Soon, photonics technologies will be used to develop optical interposers—think optical network on a chip to enhance GPU networks and memory disaggregation. The next application would be photonic computer chips.

GF has developed GF Fotonix technology to allow fabless photonics design companies to develop extremely high-speed, low-power, and low-latency photonic interconnects, connecting cutting-edge AI chips together to form large clusters that operate in unison.

What are GF Fotonix chips? 

Gupta:  GF Fotonix is GF’s third-generation silicon photonics technology. GF Fotonix is unique in the industry in that it monolithically combines active and passive photonics devices along with a 45nm class RFCMOS technology on a 300-mm SOI wafer. 

In addition to the photonic and electronic features, the technology includes features for 2D/2.5D integration (like copper pillars and copper receive pads) and “light in and light out”—v-groove-based fiber couplers with capability for on-die InP laser attach.

Silicon photonics chips using the GF Fotonix technology are fabricated in our most advanced fabrication facility in Malta, New York, along with our FinFET and other lead technologies.
 

What sets GF Fotonix apart from other chips on the market? 

Gupta: What sets GF Fotonix apart from other silicon photonics foundry technologies is the monolithic integration of photonic and electronic components on a single piece of silicon. However, there is more to this differentiation.

GF Fotonix is a feature-rich and highly flexible silicon photonics technology. We have mimicked the microelectronics model of engagement by creating a photonic technology supported by a well-characterized suite of electronic and photonic devices available in an electro-optical process design kit (PDK). This approach provides our customers with a huge time-to-market advantage.

We have customers who have designed coarse wavelength division multiplexing (CWDM) optical transceivers, dense wavelength division multiplexing (dWDM) optical transceivers, and frequency-modulated continuous wave (FMCW) LiDAR systems with optical phase arrays and little to no customization of the technology. We have also initiated work on standardizing the packaging technology.

What challenges did you and your colleagues face and overcome while designing the chips?

Gupta: While there has been significant progress in the ability to simulate photonic integrated chips, the available electronic design automation (EDA) software for photonic simulations is relatively immature compared to EDA software available to simulate electronic integrated chips. Photonic simulations, just like electronic simulations, tend to be multi-physics problems (electronic, thermal, mechanical) with the added complexity of photonic physics. GF has been working with EDA vendor partners to solve these challenges.

Evolution of optics and co-packaging technologies for data centers. 

A unique requirement for silicon photonics packaging is the need to get light in and off the PIC. Optical fibers serve this purpose. A typical SMF28 single-core, single-mode fiber consists of a ~8.2-μm core (through which the light travels) with a ~125-μm cladding. The mode field diameter is ~10.4 μm. Coupling features on the photonic IC are around 65x smaller in size than this mode field diameter. GF engineers have overcome the challenge of interfacing a large optical fiber output with microscopic features on the photonic IC with very low insertion loss.

Which applications could benefit most from the GF Fotonix chips?

Gupta: Three areas would most benefit from using GF Fotonix: optical interconnects, compute, and sensing.

GF Fotonix supports near-packaged, co-packaged, and pluggable form factors for optical transceivers. As data rates in data centers continue to increase, copper as a transmission medium has several limitations. To overcome these limitations, data centers are switching to optical fibers. GF Fotonix-based chips are being used to convert electrons to photons to transmit over optical fibers.

GF Fotonix is also being used in two areas of compute: photonic compute and optical interposers. Finally, GF Fotonix is useful for automotive sensing applications such as FMCW LiDAR. Future applications of GF Fotonix will also extend into the medical field for applications such as viral and bacterial diagnosis.

How do you envision the short- and long-term future of silicon photonics?

Gupta: The future of silicon photonics is bright (pun intended). Silicon photonics stands out as the only technology that allows extremely high levels of photonic components (and electronic components in the case of GF Fotonix) integration compared to competing technologies such as indium phosphate and silicon nitride.

Optimizing GF Fotonix Bandwidth

Gupta and his team at GF plan to continue improving the GF Fotonix platform to better serve the immediate AI/ML market. In the next few years, they hope to increase their chips’ bandwidth from 100 Gigabits per wavelength (G/λ) to 200 G/λ, and eventually to 400 G/λ and beyond.

They will also work on increasing the beachfront density of fiber interconnects and heterogeneously integrating various materials, including indium phosphide (InP) and barium titanate (BTO). Concurrently, they are trying to enhance the repairability of interconnects, using detachable and socketized optical interfaces.

All images used courtesy of GlobalFoundries.