With the rapid growth of artificial intelligence models, an unexpected alternative is emerging to the supply of high-bandwidth memory (HBM), which has become the biggest bottleneck in the sector. Kepler Computing, which has been working behind closed doors for seven years, announced that it has developed a cost-effective alternative to classic HBM solutions using ferroelectric memory (FeRAM) technology and a new 3D stacking architecture.

The most striking point is that this production can be carried out within the limits of existing and mature manufacturing, without the need for the most advanced semiconductor nodes such as extreme ultraviolet (EUV) lithography.

EUV Requirement Eliminated: FeRAM Production on a 28nm Node

Currently, memory manufacturers such as SK Hynix, Samsung, and Micron, which produce HBM, are dependent on high-cost EUV lithography systems and highly complex packaging processes. Kepler Computing aims to significantly reduce production costs while maintaining standard HBM capacities by customizing ferroelectric RAM (FeRAM) chemistry. The company states that it conducted trials of 35 different composite materials before reaching a scalable final design.

GlobalFoundries was the most critical partner in the company’s production strategy. Kepler managed to transform a standard 28nm production facility into a memory factory in a record time of just eight months. Considering that the commissioning time for classic DRAM factories is around 24 months, this conversion time provides a valuable advantage for the industry.

Investment Support and Roadmap

Kepler Computing has raised approximately $468 million in funding from investors including industry giants such as GlobalFoundries, Intel Capital and AMD Ventures in order to bring its developed architecture to life. The company’s current situation and future plans are as follows:

• Number of Wafers Processed: Approximately 2,000 silicon wafers have been processed so far using the new technology.
• First Samples: The first FeRAM-based HBM-like memory samples are planned to be delivered by the end of this year for error-free testing.
• Mass Production Date: Full-scale volume production is targeted to begin next year at GlobalFoundries’ Singapore facility, with production at the US facility expected to be operational in 2028.

Industry Skepticism and Real-World Expectations

While many memory ventures in the past have emerged with similar claims, they have struggled at the commercialization stage due to issues of yield and reliability in mass production. While FeRAM technology is known for its power efficiency and speed, it remains to be seen how it will practically meet the write endurance and massive memory bandwidth required by high-density AI accelerators.

If Kepler Computing can validate its promised performance metrics in the field, it seems possible for giants like AMD and Intel to significantly reduce HBM costs in accelerator chips and lighten the packaging burden on TSMC. Concrete information regarding the density, bandwidth, and latency of the company’s first test chips is expected to become clear in the coming months.