Earlier this week, Samsung announced that it would begin ramping up production of its 8GB HBM2 technology in order to address the rapidly expanding need for high-bandwidth memory in the market. HBM2 is used in high performance processors used for artificial intelligence, HPC (high performance compute), graphics, gaming, networking, and enterprise servers because of its low power, high throughput capability. HBM technology uses stacked memory dies that are vertically connected by TSVs (through silicon vias) to improve density and capacity, while also preventing overheating to guarantee higher reliability.
First brought to market in June of 2016, HBM2 from Samsung in an 8GB configuration is able to generate 256 GB/s of bandwidth per stack. This is an 8x increase of current GDDR5 memory technology and double the 4GB capacity of previous HBM solutions. By integrating multiple stacks with each GPU or compute processor, systems can be designed without the limitations of capacity, offering upgrade paths for graphics and data-intensive HPC workloads.
Both NVIDIA and AMD utilize HBM2 on current and upcoming graphics and AI processors, including NVIDIA’s flagship GV100 product announced in May. Though the volume of NVIDIA AI and machine learning products is low, AMD will depend on HBM2 memory for both its currently available Radeon Vega Frontier Edition professional graphics cards as well as the upcoming Radeon RX Vega product family aimed at the much larger gaming market. It has long been known that the current cost of 8GB HBM2 stacks were prohibitive to AMD being able to sell a product into the price sensitive gaming ecosystem, forcing AMD to overcharge based on the product performance or to cut margin and remain competitive with NVIDIA GeForce offerings.
Ramping up of HBM2 production, particularly of the 8GB stacks that AMD Radeon Vega utilizes, will lower costs and allow AMD more flexibility with product design and pricing. How quickly Samsung can ramp up production will be an important data point as AMD has slated the release of the high volume RX Vega family in just a few weeks. Early sales of the gaming card family may cost AMD significantly more to build than future iterations.
SK Hynix announced HBM2 production on its roadmap but the timing of its ramp of 8GB stack capacities is unknown. This will add volume to the market, in addition to competition, driving costs down further. This should give AMD more flexibility on pricing and vendor selection, improving the company’s forecast for Vega and future GPU technologies.