An artist’s view of the 4 Tbps optical I/O system demonstrated by Ayar Labs at OFC 2023. This includes two of his SuperNova light sources from the same company connected to a TeraPHY “chiplet”. [Image: Ayar Labs] [Enlarge image]
May 24, Silicon Valley-based Ayar Labs, an early developer of optical interconnect “chiplets” and light sources for high-performance computing (HPC) and AI, has increased its Series C funding by US$25 million. Announced. The new Series C1 upward round, which the firm characterized as “oversubscribed,” came from a group that included both existing and new investors. The additional cash brings Iyer’s total Series C funding to US$155 million, which began in late April 2022.
In a press release accompanying the announcement, Ayar said it would use the additional funding to “accelerate the implementation and commercialization of optical I/O.” [input/output] Especially for the market where generative AI technology promises. The company’s CEO, Charles Wuischpard, described the new funding as “a plan to further validate our technology and bring silicon photonics-based interconnect solutions to market at scale.”
Of particular interest was the participation in the Upround of Nvidia, the chip manufacturing giant behind the graphics processing units (GPUs) and other systems powering the recent surge in AI development.
Attack bandwidth and distance bottlenecks
In an interview with OPN last year, Ayar Labs co-founder and chief technology officer Mark Waid described the company’s focus as “optical computing.” In particular, Ayar has designed and built silicon photonic systems that help optically move large amounts of data flowing between application-specific integrated circuits (ASICs) in HPC and AI applications.
Replacing electronic I/O with optical I/O in these systems targets the so-called bandwidth and distance bottlenecks in HPC, AI, data centers, and other settings. In principle, using photons in fibers rather than electrons in copper wires to move as much data as possible over the longest possible distances should significantly reduce power, operating costs, and latency. In Wade’s words, his Ayar approach to achieving these benefits is to “perform the electrical-to-optical conversion as close as possible to the host ASIC.”
laser and chiplet
Ayar has two CMOS native product lines serving the HPC and AI computing markets. One is a chip-scale laser source called SuperNova. The other, called TeraPHY, is an optical I/O “chiplet” that can be flip-chip bonded to the same substrate as the electronic ASIC, right next to it.
Ayar’s TeraPHY chiplets are designed to be flip-chip bonded on the same substrate as the ASIC to simplify integration. [Image: Ayar Labs technical brief, 2021] [Enlarge image]
This architecture allows large amounts of data coming out of the ASIC to be converted to the optical domain as quickly as possible without going through copper wires. Furthermore, according to Wade, the micro-ring resonator system, which handles the conversion of electricity to light within the chiplet, provides a “much denser starting point and ultimately much more energy” than traditional pluggable transceivers. It is said that highly efficient conversion is possible.
At the 2023 OFC conference in March of this year, Iyer demonstrated what he called “the industry’s first 4 terabits per second bi-directional wavelength division multiplexed optical solution.” The demo system transports 64 wavelengths multiplexed from one of Ayar’s SuperNova light sources over eight fiber links between two TeraPHY chiplets at 2.048 Tbps in each direction and 4.096 Tbps of bi-directional bandwidth. I was allowed to. Data transfers are also very energy efficient, requiring less than 5 pJ per bit, Ayar said.
Nvidia elements
The recently announced US$25 million Series C1 Up Round will be led by new participant investment group Capital TEN, with new and existing investors and strategies including Nvidia, Applied Ventures, Global Foundries HP and Lockheed Martin. It also includes commercial partners.
The funding from Nvidia looks particularly interesting and timely, as it also participated in last year’s first US$130 million Series C funding.
The funding from Nvidia looks particularly interesting and timely, as it also participated in last year’s first US$130 million Series C funding. As widely reported, on May 24 (the same day as Mr. Iyer’s announcement), the U.S. chipmaker expects sales to reach $11 billion in the July quarter, beating analysts’ forecasts. announced more than 50%. The company’s share price surged about 27% on the day of the news, adding NVIDIA’s market capitalization to about US$184 billion, which is said to be the largest one-day gain in market history.
Behind the projected revenue growth and soaring market value is Nvidia’s pole position as a supplier of high-end chips in the burgeoning generative AI space. The company’s powerful H100 processor, his $40,000 chip specifically targeting generative AI applications, has been selling at breakneck pace since the launch of his ChatGPT in OpenAI in November 2022. increase.
Regarding the much talked about OpenAI chatbot, Nvidia CEO Jensen Huang recently said: financial times, especially among cloud computing providers such as Amazon and Microsoft, “immediate demand has been created” for the H100 processor.recently wall street journal At the event, Elon Musk, who is launching his own generative AI startup called X.ai, quipped that Nvidia GPUs are “much harder to come by than drugs.”
“Generative AI models with trillions of parameters are driving demand for this platform,” said Craig Thompson, vice president of business development for Nvidia’s Networking business unit. And that, he argued, “is why we are increasing our investment in Ayar Labs.”
