Source URL: https://www.theregister.com/2024/08/28/broadcom_optics_gpus/
Source: The Register
Title: Copper’s reach is shrinking so Broadcom is strapping optics directly to GPUs
Feedly Summary: What good is going fast if you can’t get past the next rack?
In modern AI systems, using PCIe to stitch together accelerators is already too slow. Nvidia and AMD use specialized interconnects like NVLink and Infinity Fabric for this reason – but at the 900-plus GB/sec these links push, copper will only carry you so far.…
AI Summary and Description: Yes
**Summary:** The text discusses innovations in optical interconnects and co-packaging technologies in GPU systems aimed at overcoming bandwidth limitations posed by traditional copper connections. Broadcom’s advancements could significantly impact high-performance computing and data center architectures, enabling large-scale systems with superior connectivity.
**Detailed Description:**
– The text highlights the limitations of traditional copper connections in AI systems, where high data rates surpass copper’s effective range.
– **Key Innovations:**
– **Optical Interconnects:** Moving away from copper to optical interconnects offers higher bandwidth (900+ GB/sec but limited reach with copper).
– **Power Consumption:** The switch to optical interconnects necessitates increased power (an estimated additional 20 kilowatts per rack) but could lower per-port power consumption significantly.
– **Co-packaged Optics (CPO):** Broadcom is exploring integrating optical engines directly into GPUs, enhancing bandwidth to 1.6TB/sec per chiplet while reducing power consumption to as low as five watts per port.
– **Technological Comparisons:**
– Broadcom’s 51.2Tbit/sec CPO switch versus Nvidia’s upcoming Blackwell generation, which targets 1.8TB/sec using copper.
– Previous achievements in optical chiplets by Intel and other startups, indicating a broader trend in the industry towards optimizing chip interconnects.
– **Architecture Insights:**
– The emerging co-packaged optical technology can enable scale-up systems, integrating hundreds of GPUs acting as a single large accelerator—a departure from traditional scale-out architectures.
– Broadcom’s experimental setup with test chips illustrates the potential for greater integration in data center environments.
– **Industry Context:**
– Ongoing developments by major players (like Intel, Amazon, Google, and AMD) indicate competitive advancements in optical technology for enhancing data center performance and capacity.
– Multiple startups are also contributing to the ecosystem, indicating a rapidly evolving field focused on optical solutions for high-speed computing.
This content is crucial for professionals concerned with the future of computing architectures, as it underlines significant shifts in hardware design aimed at addressing both performance and energy efficiency challenges in large-scale AI and cloud computing implementations. The insights on co-packaged optics suggest a transformative direction in how high-performance systems will communicate, ultimately influencing both infrastructure security and operational efficiency in data centers.