As the hybrid memory cube architecture matures and the enabling ecosystem develops, it follows that the form factors incorporating HMC will proliferate—and they have, thanks to the development efforts of innovative organizations like Pico Computing.
We here at Micron have been working with Pico on HMC projects for some time, so we couldn’t have been more pleased when Pico joined the Micron family earlier this year. Since then, our Micron Seattle team (as they quickly became known) has continued to help drive HMC adoption with a portfolio of HMC-based solutions. The most recent is the SB-850, a single high-performance UltraScale FPGA driving up to eight HMC devices (independent or chained), meaning no more compromises between high bandwidth and high capacity. The result is a single-board supercomputer/blade server (built on a full-height, GPU-length PCI Express platform) that accelerates myriad HPC applications, from bioinformatics to medical imaging to 8K video. Pretty cool!
The fully integrated HMC controller IP (residing on the SB-850’s FPGA) unlocks the HMC’s power, providing a whole new class of benefits to memory-bound applications, particularly those that require high bandwidth and fast random access.
The board also comes with support for OpenCL, which enables significantly faster time to market by facilitating a higher level of design abstraction. This means that software developers can easily create highly parallelized, computationally intensive applications in a familiar C-like language, without extensive knowledge of FPGA design principles. And by transferring the program execution to dedicated programmable hardware—and leveraging HMC to break through the memory wall—applications running on the Pico-850 platform will achieve dramatically improved performance and, at the same time, break the cycle of upward-spiraling CPU power costs.
In other words, for the first time, high-performance FPGA technology is available to the masses. And now, HMC is, too.