Samsung has begun mass production of 20nm second-generation High Bandwidth Memory (HBM2), which features up to 256 gigabytes-per-second (GB/sec) of available bandwidth per memory stack—double that of the first generation HBM used in AMD's Fury graphics cards.
HBM2 will also allow for HBM-equipped graphics cards to be packed with more memory—as much as 16GB—surpassing the 4GB limit that AMD ran into with its early adoption of the technology.
Like HBM, HBM2 is a form of stacked memory, where the individual DRAM chips are placed on top of each other, rather than side by side. Those chips are connected together vertically using through-silicon vias (TSVs)—wires that are threaded through the DRAM stack—while an interposer at the bottom of the stack routes the connections from the memory directly to the GPU. Because the chips are closer together and the interconnects are shorter, throughput is increased and power consumption is reduced.
Read 7 remaining paragraphs | Comments
No comments:
Post a Comment