Samsung has announced it will be increasing the production and supply of its 8GB HBM2 (second generation High Bandwidth Memory) chips in order to “address rapidly growing market demand.” Samsung isn’t providing AMD with HBM2 chips for its upcoming Radeon RX Vega, suggesting Nvidia has upped the demand for HBM2 memory for use in its next-generation graphics cards.
"By increasing production of the industry's only 8GB HBM2 solution now available, we are aiming to ensure that global IT system manufacturers have sufficient supply for timely development of new and upgraded systems,” said Jaesoo Han, executive VP of memory sales and marketing at Samsung. “We will continue to deliver more advanced HBM2 line-ups, while closely cooperating with our global IT customers".
Samsung expects 8GB HBM2 to cover half of all its HBM2 production throughout the first half of 2018, indicating 8-Hi stacks are becoming the priority over 4GB HBM2 modules.
Samsung’s 8GB HBM2 stacks are formed by eight 8-gigabit HBM2 dies and a buffer die that sits at the bottom of the stock. All eight layers are vertically interconnected by TSVs (Through-Silicon Via), with each Samsung 8GB HBM2 package containing more than 40,000 TSVs. This improves data transfer speeds by passing the connection directly through the silicon wafer rather than wire bonds.
The 8GB HBM2 stacks offer 256GB/s memory bandwidth per package. From a gaming point of view, it should help with high-end computing and rendering at higher resolutions.
The next major question is what Nvidia wants these HBM2 chips for. They've got one of two options - a Pascal refresh with marginally higher clock speeds, some optimisations, and support for HBM2, or going head first with its next-gen Volta GPU. Various rumours suggest Volta is arriving either late this year or early 2018, so its launch could well tie in with Samsung's production increase.