11 insights from the CASPA CEO Summit: AI infrastructure hardware becomes a systems game
Faster chips alone no longer define the AI buildout. Hundreds of thousands of processors now work together as one machine, so AI infrastructure hardware is designed and judged as a complete system. That system runs from the optical links between racks to the memory, storage and power that keep GPUs busy. The next phase will […] The post 11 insights from the CASPA CEO Summit: AI infrastructure…
In a recent summit marking the 35th anniversary of the Chinese American Semiconductor Professional Association, industry experts shed light on the evolving landscape of AI infrastructure hardware. Irving Tan, CEO of Western Digital Corp., emphasized that data is the heart of AI, and that storage must be strategically tiered across the cloud, edge, and devices.
Semtech Corp. supplies chips for optical transceivers and copper cables, with connectivity playing a crucial role in GPU uptime and revenue. Hyperscalers are seeking higher bandwidth at lower power, making performance per dollar per watt a key metric for inference. Rebellions Inc. designs flexible hardware to adapt to changing workloads, and Rebellions Inc. sees the network as the next bottleneck as context windows grow.
Advanced packaging, combining logic, memory, and I/O chips, is becoming the industry's driving force as Moore's Law reaches its limit. The memory wall, currently unsolved, may persist for five years, impacting models with massive parameters. Compute Express Link, or CXL, enables multiple servers to share a pool of memory, and hyperscalers are reviving interest in this technology by reusing older DDR4 modules.
New memory fabs take three to four years to become operational, leaving memory from top suppliers sold out through 2028. Unconventional AI Inc. is developing a fully integrated machine that leverages the physics of the real world to compute, targeting human-level intelligence at 10,000 watts in a single rack. Power infrastructure challenges arise as AI hardware turns over frequently, with power limits requiring operators to link multiple data centers using fiber and converting electrical connections to optical ones.
Written by urgent.news from SiliconANGLE's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
This story
This is one outlet's version. Read the fullest account.
