AMD Chair and CEO Lisa Su met with Samsung Electronics executives in Seoul on October 7 to discuss the supply of high-bandwidth memory (HBM), a component increasingly in demand for AI accelerators. Samsung wants to use that need to expand its relationship with AMD beyond memory and win new contracts for its semiconductor manufacturing business, although no agreement confirming a shift in production has been announced so far.
AMD and Samsung in 30 seconds
- Lisa Su visited Samsung’s headquarters in Seoul to discuss a broader collaboration and AI memory supply.
- The AMD Instinct MI455X accelerator uses 432 GB of HBM4, 50% more than the 288 GB of HBM3E in the MI355X.
- Samsung and SK hynix are key partners for AMD’s memory supply, which the company acknowledges remains limited.
- Samsung is looking to expand the relationship into chip manufacturing through Samsung Foundry, its foundry division.
- AMD currently relies on TSMC to manufacture its advanced chips, so adding another supplier would require validating processes, performance, and timelines.
The meeting comes as semiconductor makers try to keep pace with growing demand for AI infrastructure. For AMD, having enough accelerators doesn’t depend solely on the ability to design and manufacture the chips: it also needs to secure the memory supply that feeds its processors with data at high speed.
According to South Korean outlet ChosunBiz, the meeting brought Su together with executives from Samsung’s semiconductor division. The executive told reporters that the two companies would discuss a broader collaboration. The talks come seven months after a previous visit and at a time when AMD is looking to plan its supply chain several years in advance.
Samsung, for its part, is interested in turning its position as a memory supplier into a broader industrial relationship. Contract chip manufacturing, known as foundry, would let the company bid for new AMD business. However, Lisa Su’s public statements suggest the companies are still exploring opportunities, not that they have closed a deal to manufacture AMD’s processors or GPUs at the South Korean facilities.
HBM4 memory multiplies AMD’s supply needs
HBM memory, short for High Bandwidth Memory, is designed to deliver high bandwidth in systems that need to move large amounts of data. It is used alongside AI accelerators and other high-performance processors because it can feed their compute units with data without relying solely on conventional memory.
The capacity installed in each accelerator determines the total volume the manufacturer needs to purchase. In AMD’s case, the Instinct MI455X packs 432 GB of HBM4 memory, compared with the 288 GB of HBM3E in the MI355X. The difference amounts to a 50% increase per unit.
That means that if AMD sold the same number of accelerators from both generations, it would need 50% more memory capacity to equip them. If it also increases the number of units produced, total demand would grow even further. That’s why securing supply has become an important issue for meeting launch plans and serving data center customers.
During her visit to South Korea, Su noted that memory remains a limited resource. AMD is working with its partners to increase production and make more efficient use of available memory, at a time when Samsung has reportedly locked in most of its own memory capacity through long-term supply deals. The company also aims to plan its supply chain with a three-to-five-year outlook, rather than negotiating only the immediate needs of each product.
The executive identified Samsung and SK hynix as key partners for memory. She also said AMD had begun shipping its new-generation MI455 and Helios systems, which use HBM4. Her remarks did not detail supply volumes from each vendor or how production is split between the two companies.
The shortage isn’t limited to HBM either. According to the market analysis cited by ChosunBiz, manufacturers must split advanced production capacity between high-bandwidth memory and conventional DRAM. Boosting process yields and ramping up production of new products takes time, so the market could face supply constraints through 2027.
AMD’s interest in expanding the collaboration reflects that pressure. The more complex and demanding its accelerators become, the greater the need to coordinate memory supply with chip production and full-system assembly.
Samsung wants to expand beyond memory
Samsung and AMD signed a memorandum of understanding in March to expand their semiconductor collaboration, building on an earlier deal in which AMD turned to Samsung to secure HBM4 memory for its Helios AI platform. The agreement covers HBM4 supply for the Instinct MI455X accelerators and DDR5 memory for the sixth-generation EPYC server processors, code-named Venice, as well as for the Helios AI platform.
The relationship also opened the door to exploring opportunities in semiconductor manufacturing. For Samsung, landing foundry contracts with AMD would expand the commercial scope of the collaboration and tie memory supply to other components of AI systems.
Samsung Foundry competes in a market where fabless chipmakers and companies with outsourced production look for advanced capacity for their designs. AMD uses TSMC to manufacture its most advanced processors and accelerators. Changing or expanding that industrial base isn’t just a matter of moving a design to another fab: it requires adapting and validating the process, checking performance, reaching adequate production levels, and making sure timelines line up with the commercial schedule.
That work can add risk at a time when AMD is looking to increase deliveries of its data center products. A variation in manufacturing yield or a delay during validation could affect chip costs and availability. That’s why any eventual collaboration with Samsung Foundry would have to be assessed on a product-by-product and process-by-process basis.
The information published by ChosunBiz suggests Samsung is looking to tie memory supply to new foundry orders. However, there is no confirmation that AMD has agreed to manufacture its consumer GPUs, professional accelerators, or EPYC processors there. Nor has it been announced which chip families might be part of any expanded relationship.
After the meeting, Su reiterated that Samsung is an important partner in several areas and that AMD continues to look for collaboration opportunities. She also explained that the complexity of the supply chain requires holding early conversations with its partners. The executive also confirmed that she met with SK hynix representatives during her visit to South Korea.
The trip included other activities related to the local tech industry. Before visiting Samsung, Su took part in a meeting with South Korean government officials and companies linked to semiconductors and AI infrastructure. Topics discussed included setting up an AI research center in South Korea and possible infrastructure combining AMD CPUs and GPUs with neural processing units (NPUs) developed by local companies.
The talks between AMD and Samsung thus show two different needs that may partially align: AMD is looking to secure essential components for its AI systems, while Samsung aims to grow revenue from its foundry business. HBM4 availability is an immediate concern, but manufacturing processors and accelerators calls for broader technical and commercial decisions. For now, expanding the collaboration remains a possibility under discussion.
Frequently asked questions
Why does AMD need more HBM4 memory?
Its new accelerators need large amounts of high-bandwidth memory to move data quickly. The Instinct MI455X packs 432 GB of HBM4, 50% more than the 288 GB in the MI355X.
What does Samsung want from AMD?
Samsung is looking to expand the collaboration from memory supply into semiconductor manufacturing through Samsung Foundry. No agreement to produce AMD chips at its fabs has been confirmed.
Does AMD manufacture its GPUs and CPUs at Samsung?
According to information published by ChosunBiz, AMD relies on TSMC to manufacture its advanced chips. The talks with Samsung cover possible collaboration opportunities but do not confirm a change of manufacturer.
Which companies supply HBM memory to AMD?
Lisa Su identified Samsung Electronics and SK hynix as key partners for AMD’s memory supply. The company did not publicly detail how volumes are split between the two vendors.
via: biz.chosun.com
