Micron Technology, America’s one-and-only memory manufacturer, has come up with a cool and unusual new type of memory chip. Well, it’s not really a chip. It’s more like a module. It’s a cube, actually.
The company calls it a “hybrid memory cube” (HMC) and it starts out as a set of stacked die within one package. That’s not terribly unusual in itself; plenty of companies have stacked two or more silicon dice on top of each other, and the technique is especially useful for memories. But the HMCs use TSVs (through-silicon vias) to connect the chips together, a fairly cutting edge technology.
Micron didn’t stop there. The company also moved the memory controller, which is typically part of the CPU or the system logic, onto the memory cube itself. (Which is not really cube-shaped, by the way.) By putting the memory controller on the memory module, the HMC acts as a sort of shared system resource, able to arbitrate and manage multiple memory requests from multiple masters (CPUs, DMAs, etc.)
Finally, Micron’s HMC has its own special serial interface, which is both fast and low-power. The high-speed serial interface also keeps the pin count low; lower than a big parallel interface would be, anyway. The trouble with any new interface, however, is that it needs to be supported by the chips on the other end. So far, Micron is the only company using this interface but one assumes that they’ve been lobbying other chip makers to adopt it as well.
HMC’s are intended for high-performance systems that need a lot of memory and a lot of bandwidth. Web servers and network storage come to mind. If these news HMCs catch on, they could be a designer’s best way to pack fast, dense memory into embedded systems.