In part one of this two-part series, Steven Woo, Rambus fellow and distinguished inventor, speaks with Ed Sperling of Semiconductor Engineering about the basic design tradeoffs between GDDR6 and HBM2.
To illustrate the differences between the two memory types, Woo highlights a 256 gigabyte per second memory system built using either GDDR6 or HBM2. In part two of this two-part series, Woo and Sperling discuss how to best match a system with the most appropriate memory and explore specific applications and markets for each memory type.
HBM2 or GDDR6?
Engineers choosing between HBM2 and GDDR6, says Woo, should decide what they are trying to achieve and think about what the product goals are. No matter the specific goal, every component should be looked at holistically from a system design perspective.
“As engineers, we need to transition our thinking more to the system level. We should be taking multiple factors into account,” he elaborates. “For example, power efficiency is now being thought of as well – especially in data centers. Power consumption affects the TCO, so the decision of what memory to use really depends on the type of job you’re trying to do and how important the power efficiency is in your equation for overall TCO.”
HBM2, says Woo, works well for systems that are the most power constrained systems that require the highest levels of bandwidth – and are also willing to pay the cost and complexity to get there.
“It can be prohibitive to design this kind of system. So, I think if the costs were equal and engineering difficulty were equal to GDDR6 you would see lots of people going in this direction,” he explains. “However, because they are not equal, there is a large fraction of the market that will stay with something like GDDR6 simply because it’s easier to engineer.”
Indeed, HBM2 memory is intricately stacked, which adds cost and complexity to systems and can reduce the yield as well.
“So again, there’s additional costs here and you just have to make sure your system can actually bear the additional cost and the additional complexity that are associated with HBM2,” he adds.
Data Center Versus Consumer Applications
According to Woo, HBM2 is currently a great lead-in application for the data center, as the environment tends to be stable and controlled. However, says Woo, HBM2 will ultimately find its way into additional markets and applications as the industry becomes more experienced with manufacturing and implementing high-bandwidth memory.
“Today, the consumer market is really kind of outside the range of HBM. This is partly because most standard consumer devices don’t need the kind of performance HBM provides, and partly because of the complexity of putting the system together.”
On the Edge with HBM2 and GDDR6
Woo also says that he expects to see both HBM2 and GDDR6 find their respective ways into edge devices. “The definition is starting to really bifurcate or split into more parts. For example, there is the near edge and the far edge,” he elaborates. “As you get closer to the data center, I would expect some of those solutions to look more like the data center. So, you may see deployment of HBM2 and GDDR6 in these scenarios. As you get further out towards the endpoints, it’ll start to look more like the endpoints. It could even be things like LPDDR.”
Interested in learning more about system design with HBM2 and GDDR6? You can check out part one of this two-part series here, our HBM2 PHY product page here and our GDDR6 PHY product page here.
Leave a Reply