AMD reveals HBM: the future of graphics RAM technology

[Administrator]

AMD has officially unveiled its next-generation graphics memory solution - HBM: high bandwidth memory. Sporting an enormous increase in throughput over existing GDDR5 technology, along with equally impressive power-efficiency plus impressive space-saving attributes, HBM is set for release on AMD's next flagship graphics cards, with the firm confirming that we should see these on sale within two months. Radeon R9 390X, any one?
We attended a conference call presentation last week, given by Joe Macri, CTO of computing and graphics at AMD. He talked about the reasons behind the development of HBM - specifically that while GPUs were becoming increasingly more powerful, the existing GDDR5 memory system wasn't scaling in line. He explained that 7gbps chips are available now, with 8gbps modules in the pipeline, but there's little future in the tech: the amount of power needed to increase memory bandwidth doesn't scale in a linear fashion - the faster GDDR5 becomes, the more power hungry it is. GPUs tend to have hard TDP (thermal design power) limits, and going forward, it doesn't make sense to funnel large amounts of power through to the memory system, when more is achieved by diverting it to the GPU core.
On a broader level, GPU performance is increasing at a rate that GDDR5 cannot match, potentially increasing the chance of memory bottlenecks. A new solution is required, and that's where HBM comes to the fore. As opposed to the GDDR5 system of individual modules soldered to the board and connected to the GPU's memory controller, HBM offers up a much more refined solution. Individual memory modules are stacked one on top of another, connected by 'through-silicon-vias' (TSVs) and separated by microbumps. A single GDDR5 chip on a 32-bit interface offers up 28GB/s of throughput. In contrast, an HBM stack is 1024 bits wide, with over 100GB/s of bandwidth (AMD partner Hynix has a more precise 128GB/s metric), achieved with a significant drop in voltage too. Efficiency is also increased by sitting both the GPU core and the HBM stacks on an 'interposer' which brings the two elements much closer together.
Read more…


More...