AMD’s High Bandwidth Memory Opens Up Performance Path and New Market Opportunities

On June 16, 2015 at the E3 show AMD rolled out its high end graphics GPU, Fiji, which features its High Bandwidth Memory (HBM) on its high end desktop graphics cards, the Radeon R9 300 series.

AMD’s top of the line GPU, Radeon Fury X has 4 GB of HBM delivering up to 512 Gbits/sec of memory bandwidth (increase of 63% over previous generation), with 4,096 stream processors and 64 compute units up to 1.05 GHz. The 28nm Fury X is liquid cooled. AMD also introduced air-cooled models, Fury with 56 compute units operating at 1 GHz and the R9 Fury Nano. All of these GPUs use HBM.

AMD is touting HBM for its improved performance and power consumption compared to GDDR5. The memory technology was developed by SK Hynix. The resulting solution offers three times the performance per watt with 94% less PCB area. This enables a smaller graphics card for the Fury nano card.

The enabling technology for HBM is the 2.5D packaging technology.  AMD and SK Hynix partnered to define and develop the first complete specification and prototype. The technology employs through silicon vias (TSVs) and micro-bumps to connect one stacked DRAM to the next. The stacked die also connect via TSV and micro-bumps to a logic die which provides the PHY interface to the GPU. Up to 4 stacks surround the GPU on an interposer.

AMD developed industry partnerships with ASE, Amkor and UMC to develop the first high volume manufacturable interposer solution. The interposer connects to the package substrate. One benefit to bringing the DRAM closer to the GPU is it reduces capacitive effects. The interposer enables finer interconnects. The bus width of each stack is 1,024. Therefore, with 4 stacks the GPU has a 4,096 bus width.  The Fury X has 4 stacks of 1GByte each for a total of 4GBytes.

The initial use of HBM is in high end desktop graphics cards. In 2016 AMD may roll out other memory configurations to cover a broader range of graphics cards.

The HBM technology is not proprietary.  It is in AMD’s interests to proliferate this technology in order to generate volumes and lower the manufacturing cost.

Semico Spin

AMD has chosen a high end desktop graphics card for the first implementation of HBM. It is a high priced solution for a niche market.  Currently there is about a 35% attach rate for graphics cards in desktop and notebook PCs. The majority of PCs use the integrated graphics in the CPU or APU (AMD’s Fusion technology). AMD has about 25% share of the discrete graphics market. HBM could enable AMD to regain market share from Nvidia.

AMD presented HBM a few weeks before rolling out the graphics card implementation. At the time of HBM’s introduction the memory stacks on the interposer indicate that they could be connected to a GPU, CPU or SoC.  However, the high end desktop graphics could be the proving ground for HBM. As manufacturing costs fall other markets will open up. Semico would expect to see HBM used in notebook PCs. It could also be applied to APUs.  AMD has said that it wants to focus more on high end solutions. The APUs are used in servers as well as PCs.  HBM could also be scaled for embedded applications such as for the Internet of Things, IoT.