According to Black, the HBM design was “very large in 2012,” but yields were high even then, over 90 percent. Since then, AMD has worked to lower the manufacturing cost, which “came in line in 2014”, he said.
Because of that, Black sounded supremely confident that the HBM technology built into Fiji would work. “Remember, we’ve been working with this for eight and a half years,” he said. “We’ve seen every possible reliability issue, and we’ve beat it to death.”
If there’s a catch, however, it’s that AMD appears may have locked itself in to a single memory supplier. The interface that HBM uses is an open JEDEC spec, and open to any memory supplier. But Black said that AMD has worked closely with SG Hynix to co-define not just the interface but the physical material used. “The challenge is that we’re a lot closer to these guys than before,” he said, referring to the physical proximity of the two chips. “That means that the material set they pick has to interact with our material set, or we could end up with a reliability challenge.”
Sign up for CIO Asia eNewsletters.