Much has been made of the high bandwidth memory (HBM) being used in AMD’s upcoming Radeon R9 300 series, but in this latest presentation from AMD we get the ultimate low-down on what this means from a technical and gaming standpoint.

AMD has offer us a deep look at the 2.5D memory being offered in the flagship Radeon R9 390X graphics card, expected to be unveiled in a fortnight or so. Manufactured by the folks at SK Hynix, HBM is the hyper-fast successor to GDDR5, massively improving not just memory bandwidth but efficiency and power consumption. Rough estimates put HBM at three times faster than the current industry standard - GDDR5.


The need to move to HBM stems from the limits of GDDR5. That is to say, it's reached its limit. GDDR5 modules have a relatively large footprint and they aren't getting any smaller, coupled with a need for a large number of modules per chip. Each RAM chip  must connect up with the GPU, which means the more memory that's added, the more complex and sprawling the circuit board becomes. This in turn pushes power consumption up, makes graphics cards larger, and limits the bandwidth realistically available.




Both AMD now, and Nvidia next year with Pascal, have been forced to move onto new technology, and that tech is HBM. In the diagram below you can see how the HBM stacks up vertically, taking up a far smaller area in comparison to the R9 290X chip in the diagram above. Rather than have a host of circuit traces running to the GPU, the HBM and the GPU are directly connected through the interposer itself.
What this means is stacks of memory all capable of communicating with the GPU at lightning fast speed. Each stack, which in the case of the 390X is a four-high, connects to the GPU via a 1024-bit memory bus. A single GDDR5 memory chip with a 32-bit memory bus is capable of just 7 GB/s bandwidth, whereas a stack of HBM can theoretically achieved 125 GB/s.




The next crucial area is power efficiency. As GDDR5 has got larger and larger, its use has become unwieldy. In terms of power efficiency, HBM draws on roughly three times the bandwidth per watt consumed. This isn't just about saving on gamers' electricity bills either. Lower wattage equates to less heat, which translates to the graphics card running faster while clocked higher.


We've seen already the effect this has had on AMD's upcoming R9 390X. The leaked images so far indicate a far shorter PCB than we've come to expect from a flagship card. The combined GPU and HBM DRAM stack on a single interposer means the size of cards will be drastically reduced across the board, meaning hugely powerful mini PCs could become a household norm.


AMD will be rolling out its first graphics cards incorporating HBM in just a few short weeks. For so long trailing in Nvidia's wake, the release will give AMD a full year's head start over Nvidia's Pascal GPU, also featuring 2.5D stacking memory. This obviously hasn't gone unnoticed by AMD's corporate vice president Joe Macri, who said “Nvidia creates PowerPoints and talks in advance like they are the wonderful leader of everything. While they’re talking, we’re working.”

In a move clearly designed to hit out at Nvidia's closed doors approach, Macri also said "We do internal development with partners, we then take that development to open standards bodies and we open it up to the world. This is how Nvidia got GDDR3 and how they got GDDR5. This is what AMD does. We truly believe building islands is not as good as building continents.”

The specifics of the upcoming Radeon R9 300 series has yet to be revealed, but AMD is playing a full showcase and launch sometime in June. Sources have suggested this will either be at Computex Taipei in early June, or during E3 2015, so keep your eyes peeled on Game Debate for the latest updates.

Could AMD's year-long headstart prove crucial?