Nineteen months have passed since AMD introduced its Fiji-based R9 Fury X graphics card. With bite-size and double-wide siblings along for the ride, the Furies remain the only consumer graphics cards with HBM memory on the market. We don't know anything about how graphics cards with AMD's next-generation Vega GPUs on board will look, but some shifts in the memory market might provide one clue.
The rumor-mongering internet detectives over at WCCFTech noticed a subtle change to SK Hynix's product catalog recently, and the site thinks the change is related to the specs of AMD's next-generation graphics chip. According to the article, the last version of SK Hynix's product catalog, published in the fourth quarter of 2016, listed both 1.6Gbps and 2Gbps 4GB HBM2 modules for same-quarter availability. The current version of the book lists only the 1.6Gbps 4GB part for availability in this quarter.
AMD has not provided information about many aspects of its upcoming Vega-based products, including the capacity and layout of the memory on board, but SK Hynix's catalog notes that each 1.6Gbps stack offers 204.8 GB/s of aggregate memory bandwidth. If we assume that some of the first consumer Vega cards use 8GB of HBM2 memory, as our experience suggests they will, a pair of those 4GB stacks would be required. That configuration would provide a hypothetical graphics card with a total of 410 GB/s of memory bandwidth.
For comparison's sake, the four HBM stacks in the R9 Fury X provided 512 GB/s of peak bandwith, Nvidia's high-end GeForce GTX 1080 can access its GDDR5X memory at a rate of 320 GB/s, and the Pascal Titan X's 12GB of GDDR5X memory has a staggering aggregate bandwidth of 480 GB/s on that card's 384-bit bus. If the highest-end Vega card does use a pair of these 1.6Gbps modules, the cards would lose the bandwidth crown to Nvidia's most expensive desktop offering. Given that Fiji's HBM didn't seem to be the limiting factor in that card's performance, however, such a change might not be worth fretting over all that much—especially considering that Vega promises finer-grained data movement in its memory hierarchy than past AMD cards.
In short, it's worth taking this news with a huge grain of salt. As we noted, AMD hasn't confirmed the capacity, layout, or source of Vega's memory yet. SK Hynix might not be making its fastest HBM2 available to all comers right now, or AMD could be using HBM2 modules from a manufacturer other than SK Hynix to achieve the performance it wants. Until more details arrive about Vega, we just don't know.