Image Image Image Image Image Image Image Image Image Image

| June 24, 2017

Scroll to top

Top

No Comments

New Hybrid Memory Cube has 320GBps Bandwidth

New Hybrid Memory Cube has 320GBps Bandwidth
Charles Chong

No, you did not read it incorrectly.

Yes, the memory bandwidth actually is 320 gigaBYTES per second.

 

The Hybrid Memory Cube Consortium (HMC) is backed by 100 tech companies, with Micron, Samsung, and Hynix taking the lead in development efforts to create the next standard for DRAM, and more specifically three-dimensional DRAM.  The reason it is called three-dimensional DRAM is due to multiple volatile memory dies stacked on top of a DRAM controller.  Traditionally, these memory dies were not able to be stacked due to problems with passing an electric wire vertically through the silicon.  In order to overcome this problem, a new silicon was developed, the VIA (Vertical Interconnect Access).  What resulted was an interface that was 15 times as fast as standard DRAM while reducing power consumption by 70%.

The first iteration of the Hybrid Memory Cube will come in 2GB and 4GB flavors with an aggregate bi-directional bandwidth of up to 160GBps.  At present, the fastest DRAMs are only capable of 12.8GBps, with the large majority of current DRAM sitting at 11GBps.  When the Hybrid Memory Cube is “complete”, it will feature speeds of up to 320GBps.  In addition, HMC has announced that in 2014, the next generation of the Hybrid Memory Cube will nearly double individual data rate speeds from the current 10 to 15Gbps to 28Gbps.

In the words of JD from Scrubs: ZOOM ZOOM ZOOM*!

 

 

*Yes, I realize “ZOOM ZOOM ZOOM” is used after a retort in Scrubs.  I want to use it anyway.

Tags
2 comments
ATIsux
ATIsux

So how long until we receive such upgrades on current Nvidia Architecture?

Charles Chong
Charles Chong

@ATIsux Once this technology is more widely available, I'm positive we will begin to see this in GPUs as well.  As of now, it's still a prototype, so most likely it will be implemented in actual products sometime next year.  GPUs will then (I'm guesstimating here) catch up soon after.  I'd give it 2-3 more years before it's fully realized in all our computing hardware.