JOHN WILLIAMSON AI/ML AND PHOTONICS
A PARTNERSHIP MADE IN HEAVEN? The meshing of Artificial Intelligence/Machine Learning (AI/ML) and Photonic Integrated Circuits (PICs) looks to be a productive courtship currently being made in technology heaven, writes John Williamson . Accordingly, a recent analysis from Coherent Market Insights (CMI), valued the 2022 global silicon photonics market at US$1,584.8 million growing to US$8,317.9 million by 2030. CMI believes two of the main drivers of this are the increasing adoption of SiPho in data centres and the not-unrelated integration of SiPho with AI and ML. AI/ML AND PHOTONICS:
A I/ML has a ravenous resources, and lower latency – all of which PICs could help furnish in spades. “The all-to-all connectivity requirements (of) AI workloads demand higher bandwidth density and lower cost than optics used in traditional front-end compute networks”, instances Manish Mehta, VP of Marketing and Operations, Optical Systems Division at Broadcom Inc. Christian Urricariet, senior director of Product Marketing, Silicon Photonics Product Division, Intel Corp, acknowledges that electrical I/O (i.e., copper trace connectivity) can support very high bandwidth density with low power, but only for very short distances of about 1 metre, which makes it increasingly inadequate for advanced AI/ML applications. He also allows that pluggable optical transceiver modules used in requirement for increased speed and bandwidth, higher performance computational current data centres and early AI clusters can increase the maximum reach but at cost, power density, and latency levels that cannot support the scaling requirements of the emerging AI/ML infrastructure. “Significantly scaling AI/ML network and compute infrastructure will drive exponential growth in I/O bandwidth
ARTIFICIAL SWEETENERS In reverse, AI/ML can aid in the
with more efficient resource utilisation such as xPU disaggregation and memory pooling,” remarks Urricariet. POWER POINTS Reducing AI/ML PIC power consumption is a big deal. “One of the challenges for building large AI clusters is power consumption,” states Dr. Vladimir Kozlov, founder and CEO of market intelligence house LightCounting. “Optics doesn’t consume as much.” In this context, Mehta adds that his company’s SiPho PICs, integrated with CMOS drivers and transimpedance amplifiers, and co-packaged on a common substrate with a core ASIC – for example switch ASIC or xPU - can deliver up to 70% power consumption savings versus traditional optical interconnects. Additionally, says Mehta, the solution can offer 6.4 Tbps in just two-times the silicon area of the company’s 400 Gbps PIC - an eight-times improvement in silicon area efficiency. Meantime, Intel’s first Optical Compute Interconnect (OCI) chiplet recently demonstrated, is a 4 Tbps bidirectional device, supporting 64 lanes of 32 Gbps data in each direction over 10s of metres. This is realised as eight fibre pairs each carrying eight DWDM wavelengths, and it includes the InP lasers heterogeneously integrated on the PIC. Its target energy efficiency is <3 pJ/bit, compared to current pluggable transceivers at about 12-14 pJ/bit.
germination of new PIC solutions. “The main way is laying out and designing components and circuits,” offers Dr. Adam Carter, CEO of Photonic Application-Specific Integrated Circuits (PASIC) chip producer OpenLight Photonics. Vikas Gupta, senior director of Product Management at GlobalFoundries, notes that 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 comparatively immature when 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,” says Gupta, expanding on the theme. “AI/ML can change this dynamic by making these multi-physics problems computationally more efficient and accurate.” AI could also be relevant to solving some of the well-known shortcomings and complexities of the PIC industry supply chain. ”AI could help because it could help you to model your supply chain better,” reasons Carter. “It can model component shortages and things like that.”
density and more efficient power management, coupled with longer
reaches in connectivity to support larger CPU/GPU/TPU clusters, and architectures
10
www.opticalconnectionsnews.com
| ISSUE 37 | Q2 2024
Made with FlippingBook interactive PDF creator