Hybrid Memory Cube - Take it Easy

By Susan Platt - 7.16.15

You can always spot the early adopters of new technologies; like pioneers, they’re the ones with the arrows in their backs. But like the pioneers, they brave technology’s new frontiers for the very same reasons—tremendous opportunity and huge competitive advantage. And also like the pioneers of old, they make the new territory safe for the rest of us by wringing out the technology, working out the bugs, getting it up and running—taming it, really—and ultimately demonstrating the very real and substantial rewards that made the risks so worthwhile.

But you know, it really doesn’t have to be like that. In fact, we believe in equal opportunity for all engineers—not just the early adopters—to jump in and reap the myriad benefits that the new frontier promises. And that is why our entire focus has been on domesticating and democratizing new technologies right out of the gate. No more waiting on all that “crossing the chasm” stuff. We’ve leveled the playing field.

Our breakthrough hybrid memory cube is a great example. Exotic technology? You bet. A very different memory paradigm? For sure. But hard to understand, use, evaluate, and integrate into a system? Not at all, thanks to some of our newest team members and here’s why. 

Back in January of this year, Micron acquired Pico Computing, a trailblazer in their own right.  Pico’s unique specialty was, and continues to be, FPGA hardware accelerator platforms.  These platforms offer a unique value proposition for HMC customers—one that can help them become pioneers in their own right.  Here’s how.

Take the Pico AC-510 supercomputer module as a case in point. It’s only the size of a business card, and yet it hosts a next-generation Xilinx UltraScale FPGA and a high-bandwidth HMC device optimized for random access performance. The module snaps onto a PCIe backplane, which fits into a standard PCIe slot.

Let’s deal with the FPGA first. To make things simple, we’ve created a novel framework that includes the APIs, host interfaces, and all other overhead required to load a bitfile and monitor/control performance. In other words, the framework, working with Pico’s PCIe backplane, does all the heavy lifting where programming the FPGA is concerned. Even seasoned FPGA designers will appreciate that.

The Pico AC-510’s FPGA (which, by the way, features 15G transceivers) also contains the complete HMC memory controller, enabling quite a bit of configuration flexibility. Pretty simple. No further hardware development required; just load your FPGA bitfile and you’re off and running.

Now, if creating the FPGA design is an issue (it’s no secret that FPGA design work is an esoteric, if not dark, engineering practice), you can opt for another easy path: OpenCL—the C-like language that allows software developers and integrators to focus on their algorithms without having to be concerned with the particulars of programming an FPGA. Again, pretty easy.

All told, it’s the ideal low-cost—and scalable—entry point into the world of HMC and accelerated computing. It’s scalable because as many as six Pico AC-510 modules can fit onto a single PCIe backplane, creating massively parallel computing capacity. And because those backplanes plug right into your existing PCIe environment, the system can reside anywhere from the server room (with up to eight fully populated PCIe backplanes in a single 4U rack) to your desktop. In short, we’ve made it easy so you can take it easy.

And—there are no arrows to worry about. 

Susan Platt

Susan Platt