800G/1.6T optics powering AI revolution

Anyone familiar with the movie ‘2001, A Space Odyssey’ will know that the concept of AI goes back at least 56 years, to the HAL 9000 computer which did so much to spoil the astronauts’ day. 

The writer, Arthur C Clark, claimed the acronym HAL didn’t come from shifting the letters ‘IBM’ one forward in the alphabet, as rumored, but stood for ‘Heuristically Programmed Algorithmic Computer’, which happens to be a pretty good starting point for defining an AI machine. 

A traditional computer is programmed using fixed algorithms and instructions. AI sets itself apart by combining these with heuristics, or the ability to learn for itself and make decisions or create content based on what it has learned, thus mimicking human intelligence. 

The history of artificial intelligence goes back long before the fictional HAL 9000.

• Automata, human-like dolls and machines that seem to move by their own will, existed as long ago as the 17th century, as highlighted by a CNET¹ article, ‘Before they were robots’.
• In the summer of 1955, John McCarthy, a mathematics professor and pioneer in artificial intelligence, coined the term during a workshop at Dartmouth College, and it was in vogue through the 1950s.
• Fast forward to 1966, when Joseph Weizenbaum created the first chatbot, ELIZA, inspired by Eliza Doolittle in Pygmalion. ELIZA was a simulated psychotherapist which used natural language processing (NLP) to converse with humans.
• In the 1980’s, Japan’s Fifth Generation Computer Project (FGCP) set out to create a fifth generation of computers focused on inference based processing and AI capabilities. While it is little discussed now, it represents an early recognition of the potential transformational role of AI.
• Chess grand master Gary Kasparov was famously and stunningly defeated by IBM’s Deep Blue computer in 1997.
• A few years later, NASA’s Mars Rovers Spirit and Opportunity started navigating the red planet on their own. Today, there are two Mars Rovers, Curiosity and Perseverance, active on Mars. Both are equipped with the AEGIS AI system, using deep learning, computer vision and neural networks to investigate geological targets.
• Apple Siri launched in 2011, with Amazon Alexa coming to market in 2014, and they have been helping us pick the best take-out on Friday night ever since.
   

These examples trace what was largely evolutionary development of AI, until November 30th 2022, when the launch of ChatGPT, and the first widely Generative AI capabilities, heralded what can be seen as the start of the AI revolution.

Whatever this revolution means for society, it is certain that it will have a big impact on the computing and datacenter industry, and innovations like 800G/1.6T optics are helping it rise to the challenge.

Generative AI training modules and datasets pose significant challenges in terms of large-scale data processing and low latency demands on the AI datacenter networks. Articles from Juniper Networks² and Network World³  detail how two major datacenter network providers are addressing these., while a 2023 Keysight blog discusses the vital role of 1.6T networking for emerging technologies like AI.

To meet the growing bandwidth and speed demands of AI datacenter networking, both 800G ethernet and 800G/1.6T optics are required, with 800G ethernet currently in development. Various technology options exist, from traditional Linear- drive Pluggable Optics (LPO) transceivers to Co- Packaged Optics (CPO). While each of these options has its own pros and cons, they all play a part in the overall movement to 800G/1.6T.

Priyank Shukla of Synopsys, a member of the Ethernet Alliance, predicts that AI datacenter networks will be the first adopters and are likely to deploy 800G ethernet by 2025 and 1.6T ethernet by 2027. 800G optical transceivers have been released by many vendors. and 1.6T transceivers have also been demonstrated. A January 2024 Ethernet Alliance blog⁴ discusses these and other developments in more depth.

To support the AI datacenter networks, equipment providers are bringing 800G capable 51.2Tbps switches providing AI server connectivity. These switches can support 32X OSFP800, QSFP-DD800 ports, enabling the use of 800G/1.6T optics. 800G/1.6T optical transceivers address the challenges of the AI datacenter by providing increased bandwidth and speeds to efficiently process data and reduce latency for AI applications.