JEDEC Approves SPHBM4 Standard, Aims to Lower AI Chip Costs with Cheaper Packaging

JEDEC has finalized the SPHBM4 standard, a new memory technology offering HBM4's high performance with significantly fewer connection pins.

This is crucial because memory and advanced packaging now make up over 60% of AI chip costs, and SPHBM4 is designed to make these components more affordable.

While not a replacement for top-tier packaging, it provides a vital, lower-cost alternative that can help ease production bottlenecks and broaden access to high-bandwidth memory.

000660.KS005930.KSNVDA.KSTSM.KSMU.KS

A new standard has just been approved that could significantly change the economics of building AI hardware.

This new standard is called SPHBM4, and its goal is simple: to deliver the performance of next-generation HBM4 memory without the sky-high costs associated with it today. It achieves this by using a clever serial interface, which dramatically reduces the number of physical connection pins needed from 2,048 down to just 512. Fewer pins mean the memory can be mounted on cheaper, more conventional materials called 'organic substrates' instead of always requiring large, expensive silicon 'interposers'.

This development comes at a critical time. The cost of AI chips has been exploding, driven largely by two components: the HBM memory itself and the advanced packaging needed to connect it to the main processor. In fact, by late 2025, HBM and packaging accounted for roughly 60-63% of an AI accelerator's component costs. This has created a major bottleneck, limiting how many AI chips can be produced and keeping prices high.

The path to SPHBM4 was paved by several key factors. First, recent announcements confirmed the industry's direction. TSMC, a leading chip manufacturer, stated that its high-end CoWoS packaging would remain essential for the biggest AI chips, reinforcing the need for cheaper alternatives for everything else. At the same time, memory makers like Samsung and Micron began shipping HBM4 samples, proving the technology was ready and creating demand for more flexible ways to use it.

Second, this standard addresses a long-running capacity problem. For years, the supply of advanced packaging like CoWoS has been a major constraint. By allowing some designs to move off these capacity-limited production lines, SPHBM4 can free up space for the truly massive chips that still require them.

Ultimately, SPHBM4 provides a new, vital tool for system designers. It's not a magic bullet that solves all scarcity issues, but it introduces a crucial trade-off. It allows for a more cost-effective way to achieve high memory bandwidth, giving companies more options to design a wider range of AI accelerators. This could help democratize access to high-performance AI hardware and ease the supply chain pressures that currently define the industry.

HBM (High Bandwidth Memory): A type of high-performance RAM where memory chips are stacked vertically, like a skyscraper, to provide much faster data access for powerful processors like those used in AI.
Organic Substrate: The foundational layer, similar to a mini-motherboard, on which chips are mounted. Organic substrates are a more conventional and cost-effective material compared to silicon interposers.
CoWoS (Chip-on-Wafer-on-Substrate): A high-performance 2.5D packaging technology developed by TSMC. It's essential for connecting HBM to high-end processors but is expensive and has limited production capacity.