JEDEC Finalizes HBM4 Spec With A Key Upgrade For Memory Manufacturers

With CPU core counts and GPU silicon sizes simply skyrocketing, it's become harder and harder to feed all that processing power with data. Regular old DRAM packages just aren't cutting it anymore for server systems. Even for CPUs, we're seeing a clear transition over to High Bandwidth Memory (HBM), which offers, well, you probably can guess from the name. HBM achieves extremely high bandwidth by using a super-wide memory interface that's expensive to implement in consumer hardware, but offers the speed that enterprise systems need.

HBM is much newer than the familiar DDR SDRAM and even the "GDDR" variants used in GPUs, yet it's quickly catching up to standard DDR in terms of revision number. JEDEC, the organization that creates and ratifies new memory standards, has just announced that it is very close to finalizing the HBM4 standard, improving both performance and capacity in an "evolutionary" way.


HBM4 is also improving capacity, with support for both 3GB and 4GB layers stacked in 4-hi, 8-hi, 12-hi, and 16-hi configurations. This means that a single stack of HBM4 RAM could give a processor as much as 64GB of RAM, connected by a 2048-bit interface for, hypothetically, over 2.5 TB/second of memory bandwidth—assuming such a dense stack of HBM4 could hit the same peak 9.8 Gbps transfer rates that Samsung demonstrated with HBM3e.

It would be exciting to see HBM find its way into consumer components again—the last consumer-facing part to sport stacks of HBM was the Radeon VII—but it's pretty unlikely given the insane demand for the fast RAM from the enterprise compute world. All those NVIDIA Blackwell, AMD Instinct, and Intel Data Center MAX GPUs make a whole lot more money than consumer graphics cards. Well, a gamer can dream.