[News] Unpacking AMD’s MI350: Powered by TSMC’s N3P, with Samsung/ Micron as Dual HBM3E Suppliers

AMD officially unveiled its MI350 series GPUs at the Advancing AI 2025 event on June 12—and its key suppliers are now coming into focus. According to Commercial Times, both the MI350X and MI355X are built on TSMC’s N3P node. On the memory front, Sedaily reports that the chips pack 12-high HBM3E from both Samsung and Micron, marking a big win for Samsung amid tough competition.

According to AMD, its MI350 series GPUs will enter mass production and shipments in Q3, followed by the MI400 series and its first rack-level AI solution, Helios, slated for launch in 2026. As Reuters notes, the AI chip race is shifting to full-server solutions, and AMD’s upcoming Helios server, powered by 72 MI400 chips, is designed to rival NVIDIA’s NVL72 in this new battlefield.

But for now, AMD’s MI350 series suppliers are drawing plenty of buzz. Here’s a quick look at its key partners in chipmaking and memory.

TSMC Scores Big with AMD’s MI350 and NVIDIA’s Rubin on N3P

Commercial Times suggests that AMD’s MI350 and MI355 are built on TSMC’s N3P node and use its advanced SoIC and CoWoS-S packaging—putting them a step ahead of NVIDIA’s upcoming Rubin GPUs for now.

Meanwhile, another Commercial Times report notes that NVIDIA’s Rubin will also use TSMC’s N3P but with the larger CoWoS-L packaging. Set for mass production in early 2026, Rubin is expected to drive even more demand for TSMC’s cutting-edge tech, the report adds.

Samsung’s HBM3E Lands Official Spot in AMD’s MI350 Series

On the other hand, while Micron announced its 12-high HBM3E is featured in the MI350 series, Sedaily reports that AMD has also confirmed the adoption of Samsung’s 12-high HBM3E, making both players key memory suppliers.

Notably, this is good news for Samsung, which reportedly hit a setback in its third attempt to pass NVIDIA’s HBM3E 12-layer validation in June. According to Business Post citing securities analysts, the company now aims for a retest in September.

As AMD stated in its press release, built on the AMD CDNA 4 architecture, the AMD Instinct MI350X and MI355X GPUs are purpose-built for the demands of modern AI infrastructure. These GPUs deliver leading memory capacity (288GB HBM3E from Micron and Samsung Electronics) and bandwidth (up to 8TB/s), ensuring exceptional throughput for inference and training alike, it said.

With the MI350 series debuting this year and the MI400 coming in 2026, Sedaily notes rising expectations for Samsung to supply HBM4 to AMD. However, Samsung needs to work harder to turn the tide, as both SK hynix and Micron have already begun shipping 12-high HBM4 samples to major clients.

Read more

• [News] Samsung Reportedly Stumbles Again on NVIDIA’s 12-Hi HBM3E Validation, Retest Set for September
• [News] TSMC Gears Up for SoIC, Reportedly Secures Orders from NVIDIA’s Rubin After AMD and Apple

(Photo credit: AMD)