AMD Helios AI server rack enters production with Samsung HBM4 memory supply
Seoul, July 24
Samsung Electronics is supplying high-bandwidth memory for AMD's Helios AI server rack as the system enters production to compete against Nvidia in the data center market, according to a news report by The Korea Herald.
AMD Chief Executive Officer Lisa Su announced the milestone on Thursday during the company's Advancing AI conference in Seoul. Su stated that shipments were scheduled to start late in the third quarter before expanding into the fourth quarter. Built as a complete cabinet rather than individual chips, every Helios unit integrates 72 Instinct MI455X accelerators together with central processors, networking gear, and cooling systems.
The configuration supports up to 432 gigabytes of HBM4 per accelerator, totaling approximately 31 terabytes across the rack cabinet. High-bandwidth memory stacks memory layers vertically beside an AI processor to speed up data delivery. Samsung's mass-produced HBM4 holds 36 gigabytes within a 12-layer stack, amounting to roughly twelve stacks per accelerator.
By comparison, Nvidia's competing Rubin processor uses eight stacks, meaning each shipped AMD rack carries a higher volume of Samsung memory than a comparable Nvidia setup.
Comparing performance parameters, Su said Helios offers 50 per cent more memory capacity and 6 percent higher bandwidth than Nvidia's Vera Rubin platform.
As per the news report, the figures reflect AMD's internal specifications rather than third-party benchmark evaluations. Samsung gained primary HBM4 supplier status for the MI455X under a contract signed in March at its Pyeongtaek chip facility, which Su visited during the agreement. Primary supplier status does not imply exclusivity, and neither firm disclosed pricing or volume allocation. Samsung initially began commercial HBM4 shipments in February.
During a Samsung-hosted session at the Seoul conference, Bart Walker, vice president of memory for the AMD account at Samsung Semiconductor US, and senior engineer James An explained the underlying technology. Samsung did not issue a separate press release for the presentation.
Addressing the architecture, An said the HBM4 doubles the number of input-output connections to 2,000 from 1,000 in the previous generation, and that Samsung produced the logic layer beneath the memory stack on its own 4-nanometer foundry process. Samsung also supplies server memory modules for the sixth-generation EPYC "Venice" processors used in Helios.
AMD announced hardware commitments from OpenAI, Oracle, and Anthropic, though actual deployments operate on staggered timelines extending through 2027.
— ANI
Reader Comments
Interesting times for AI hardware. Samsung's 4nm foundry for the logic layer and doubling I/O connections to 2000 seems like a real engineering feat. Hope this also helps bring down costs for enterprises in India – we need more affordable AI compute options for startups and research. 🙏
Lisa Su visiting Seoul for the announcement – strategic move considering Samsung's dominant role in memory supply chain. But I'm wondering: with OpenAI, Oracle, and Anthropic as customers, will this actually disrupt Nvidia's near-monopoly in data centers? Also, staggered deployment till 2027 seems like a long timeline. Let's see if AMD can deliver on the hype.
Good to see innovation happening, but I wish Indian semiconductor companies were also part of this supply chain. We're still heavily import-dependent for high-end chips and memory. This deal between Samsung and AMD just reinforces how far behind we are. Need more local fabs and R&D investment, yaar. 😔
From a tech perspective, 432GB HBM4 per accelerator sounds insane – that's roughly 31TB per rack! The fact they're using 12-layer stacks vs Nvidia's 8 is a major memory advantage. But reliability and thermal management will be key. Hope Samsung's yield is good because that's a lot of expensive silicon in one cabinet.
Finally some real competition for Nvidia! But I'm a bit skeptical – AMD has promised many times to challenge Jensen's company but hasn't delivered at scale. The fact that deployments won't fully happen until
We welcome thoughtful discussions from our readers. Please keep comments respectful and on-topic.