AMD eases HBM risk

AMD just removed one of the easiest ways an artificial intelligence chip launch can go wrong: running out of memory chips before customers can get the servers. On March 18, Samsung said it signed a memorandum of understanding to become a key supplier of fourth-generation high-bandwidth memory for AMD’s next Instinct MI455X accelerators. (samsung.com) High-bandwidth memory is the stack of ultra-fast memory chips that sits right next to an artificial intelligence processor, like putting a pantry beside the stove instead of across town. If that memory is late, the whole accelerator card ships late, even if the main processor is finished. (datacenterdynamics.com) AMD’s new deal points at a specific product, the Instinct MI455X, which Samsung called the target for its high-bandwidth memory fourth-generation supply. Bloomberg reported Samsung would be the primary supplier for that chip, which is aimed at corporate and cloud data centers. (bloomberg.com) This matters because AMD is no longer selling just single chips. Korea JoongAng Daily reported that AMD’s Helios rack-scale system ties together dozens of MI455X accelerators in one rack, so a memory shortage on one part can slow an entire rack rollout. (koreajoongangdaily.joins.com) Samsung was already supplying AMD with high-bandwidth memory third-generation extended chips for the current MI350X and MI355X products. The new agreement extends that relationship into the next memory generation instead of forcing AMD to scramble for capacity later. (reuters.com) That is the quiet risk in the artificial intelligence server market: the bottleneck is often not the processor design but the small set of companies that can package advanced memory at scale. Reuters said Samsung signed the AMD deal during Nvidia’s developer conference week, when Nvidia was also talking up Samsung’s high-bandwidth memory fourth-generation work. (reuters.com) The agreement goes beyond graphics processors. Samsung also said it will work on double data rate fifth-generation memory for AMD’s sixth-generation EPYC “Venice” server processors and the Helios platform, which means AMD is trying to line up memory for both the brains of the server and the accelerators attached to them. (news.samsungsemiconductor.com) There is also a second clause that investors will notice: the two companies said they will discuss a foundry partnership for future AMD products. That does not mean AMD is leaving Taiwan Semiconductor Manufacturing right now, but it does mean AMD is trying to reduce single-point dependence in another part of the supply chain too. (reuters.com) AMD is still chasing Nvidia in data-center artificial intelligence, and memory alone will not close that gap. But when a rack can contain dozens of accelerators and tens of terabytes of high-bandwidth memory, securing the memory early is the difference between a roadmap slide and a product customers can actually install. (datacenterdynamics.com)