|
Optical networking is emerging as the next major trend in artificial intelligence (AI) infrastructure as growing AI computing demand drives the need for faster data exchange, lower latency and larger GPU clusters, according to a report by Goldman Sachs.
The report stated that networking is becoming a key component in unlocking computing capability for AI chips by connecting multiple chips together and enabling seamless communication between them. "Networking is the next frontier in AI infrastructure, poised to enhance computing capability through seamless data exchange and low latency," the report said. Optical networking is a communications system that transmits data as pulses of light through hair-thin glass or plastic fibre-optic cables. It replaces traditional electrical signals over copper wires, delivering massive bandwidth and extremely high-speed. According to Goldman Sachs, the rapid expansion of AI infrastructure and increasing computing power per rack are expected to create strong growth opportunities across all networking configurations. The report estimated that the total addressable market (TAM) across scale-up and scale-out networking could increase nearly nine times to USD 154 billion by 2028 from around USD 15 billion in 2026. Goldman Sachs said scale-up networking, which involves adding more GPUs and computing resources within the same server rack or across connected racks, is expected to account for around 69 per cent of the total USD 154 billion TAM, equivalent to nearly USD 106 billion. The report also noted that Co-Packaged Optics (CPO) is expected to play a major role in future AI networking systems. According to the report, CPO could contribute around USD 91 billion, or 59 per cent of the total USD 154 billion market opportunity, assuming a 29 per cent penetration rate in scale-out networking. Co-Packaged Optics (CPO) is an advanced hardware architecture that integrates optical transceivers and electronic processing chips (like switch ASICs or AI accelerators) on the same silicon substrate. It eliminates traditional pluggable modules. The report also said dollar content in networking systems is expected to increase sharply as AI infrastructure becomes more advanced. The report also projected a 13 times larger addressable market for optical modules and optical engines as networking expands from scale-out systems to scale-up systems. It explained that the AI industry is increasingly attempting to connect larger numbers of GPUs to build bigger AI clusters and improve computing performance. The report identified three major methods of AI infrastructure expansion. The first is "scale up", where more GPUs are connected within the same equipment or across connected racks to create supernodes with optimized networking speeds. The second is "scale out", which involves connecting multiple pieces of equipment through switching technologies. According to the report, modern AI clusters now support scale-out connections of more than 100,000 GPUs. The third method is "scale across", where servers across different data centres are connected. Goldman Sachs noted that Nvidia has introduced scale-across networking solutions using its in-house Ethernet switch and network interface controller products. (ANI)
|
