Echoing the rest of the DRAM market, HBM2 memory has been in short supply. There are only two companies producing the parts: Samsung and SK Hynix. Of the two, only Samsung is currently making 8GB modules (although that could change later this year). Today, Samsung announced that it's going to step up production of 8GB HBM2 packages to serve "rapidly growing market demand."
It's not hard for a PC enthusiast to imagine where those 8GB HBM2 parts might be going. AMD is already using 8GB HBM2 packages in its Radeon Vega Frontier Edition graphics cards, as well as its close cousins in the Radeon Instinct, Radeon Pro, and Radeon RX families. For its part, Nvidia doesn't seem to have tapped 8GB packages yet—even the mighty Tesla V100 uses four 4GB packages—but that could simply be down to a lack of supply. It's easy to imagine a forthcoming Quadro or Tesla part using two, three, or four 8GB HBM2 packages.
As a brief refresher, 8GB HBM2 dies can move 256 GB/s over a 1024-bit bus. Samsung trumpets the fact that it has over 850 patents (both awarded and pending) on technologies related to HBM2 memory, and that each package has over 40,000 through-silicon via connections. The company also says that it expects 8GB HBM2 packages will comprise more than half of the company's HBM2 manufacturing by the first half of next year. This writer wonders whether that means Samsung is seeing low demand for 4GB HBM2 packages, whether supply is already sufficient for those chips, or if the company is simply content to let its competition serve those customers. Only time will tell, we suppose.