Jedec readies sphbm4 standard to match hbm4 bandwidth with fewer pins

Jedec is finalizing the sphbm4 memory standard, aiming to deliver hbm4-level throughput for artificial intelligence accelerators while cutting pin count through higher interface frequencies and serialization.

Jedec solid state technology association announced it is nearing completion of a new standard called standard package high bandwidth memory, or sphbm4, which is designed to bring hbm4-class performance to more conventional packaging. According to the association, sphbm4 devices are similar to the hbm4 devices commonly used in artificial intelligence accelerators, since they use the same dram dies on a new interface base die that can be mounted on standard organic substrates, whereas hbm4 is typically mounted on silicon substrates.

The organization stated that, as planned, sphbm operates at the same aggregate data throughput as hbm4 while using fewer pins, and it does this by operating at a higher frequency. The design goal is to keep the overall bandwidth comparable to existing hbm4 solutions while making the memory easier to integrate with mainstream packaging technologies that rely on organic substrates instead of more complex silicon interposers.

Jedec explained that where the hbm4 interface has 2048 data signals, when published, sphbm4 will define 512 data signals with 4:1 serialization to achieve the same bandwidth. The association noted that this change allows the relaxed bump pitch required for connection to organic substrates, highlighting how the reduction in data signals and the use of serialization are intended to simplify physical implementation while retaining hbm4-level throughput for demanding workloads such as artificial intelligence acceleration.

52

Impact Score

Indiana launches Artificial Intelligence business portal

Indiana is rolling out IN AI, a statewide portal meant to help employers adopt Artificial Intelligence with practical guidance, workshops and peer support. State leaders and business groups are positioning the effort as a way to raise productivity, wages and job growth while keeping workers at the center.

Goodfire launches model debugging tool for large language models

Goodfire has introduced Silico, a mechanistic interpretability platform designed to let developers inspect and adjust model behavior during development. The company is positioning it as a way to give smaller teams deeper control over open-source models and more trustworthy outputs.

Nvidia launches nemotron 3 nano omni for enterprise agents

Nvidia has introduced Nemotron 3 Nano Omni, a multimodal open model designed to support enterprise agents that reason across vision, speech and language. The launch extends Nvidia’s push beyond hardware into models and services while targeting more efficient agentic workflows.

Intel 18A-P node improves performance and efficiency

Intel plans to present new results for its 18A-P process at the VLSI 2026 Symposium, highlighting gains in performance, power efficiency, and manufacturing predictability. The updated node is positioned as a stronger option for customers seeking 18A density with better operating characteristics.

EA CEO defends broader Artificial Intelligence use in game development

EA CEO Andrew Wilson defended the company’s internal use of Artificial Intelligence after employee claims that the tools were slowing work rather than helping. He framed the technology as an aid for repetitive quality assurance tasks, even as concerns persist over its broader impact on development.

Contact Us

Got questions? Use the form to contact us.

Contact Form

Clicking next sends a verification code to your email. After verifying, you can enter your message.