Micron Technology (Nasdaq:MU) has started shipping samples of its new high-bandwidth memory product to its key select customers. The 36GB 12-high HBM4 is aimed at accelerating the development and deployment of next-generation AI platforms, particularly in data centers and cloud environments.
Built on Micron’s 1-beta DRAM process and utilizing its established packaging technology, the HBM4 memory features a 2048-bit interface and speeds exceeding 2.0 terabytes per second per memory stack. This makes Micron’s 36GB 12-high HBM4 outpace its HBM3E product by over 60% in bandwidth and 20% in power efficiency, claimed the company.
“This expanded interface facilitates rapid communication and a high-throughput design that accelerates the inference performance of large language models and chain-of-thought reasoning systems. Simply put, HBM4 will help AI accelerators respond faster and reason more effectively,” Micron said.
As AI workloads — particularly training and inference at scale — become increasingly limited by memory bandwidth and power efficiency, the availability of HBM4 will mark a pivotal moment in the AI memory landscape.
“The launch of HBM4 marks a critical milestone for generative AI and large language models (LLMs), especially as these models scale to hundreds of billions or even trillions of parameters,” said Manish Rawat, analyst, TechInsights. “At such scales, memory bandwidth and capacity become the primary performance bottlenecks, more so than compute power. HBM4 addresses these challenges by offering significantly higher bandwidth (up to 1.5 TB/s per stack) and increased memory capacity (up to 64GB or more per stack), while also improving power efficiency.”
Rawat added these enhancements enable faster model training, support larger batch sizes, and deliver more efficient inference—crucial for advanced models like GPT-5, Gemini, and Claude. HBM4 also facilitates tighter integration between AI accelerators and memory, reducing latency and boosting throughput, allowing system architects to build more optimized compute stacks, especially important for real-time and multimodal AI applications.
Race to power the next generation of AI
By shipping samples of the HMB4 to the key customers, Micron has joined SK hynix in the HBM4 race. In March this year, SK hynix shipped the 12-Layer HBM4 samples to customers. SK hynix’s HBM4 has implemented bandwidth capable of processing more than 2TB of data per second, processing data equivalent to more than 400 full-HD movies (5GB each) in a second, said the company.
“HBM competitive landscape, SK hynix has already sampled and secured approval of HBM4 12-high stack memory early Q1’2025 to NVIDIA for its next generation Rubin product line and plans to mass produce HBM4 in 2H 2025,” said Danish Faruqui, CEO, Fab Economics.
“Closely following, Micron is pending Nvidia’s tests for its latest HBM4 samples, and Micron plans to mass produce HBM4 in 1H 2026. On the other hand, the last contender, Samsung is struggling with Yield Ramp on HBM4 Technology Development stage, and so has to delay the customer samples milestones to Nvidia and other players while it earlier shared an end of 2025 milestone for mass producing HBM4.”
Faruqui noted another key differentiator among SK hynix, Micron, and Samsung: the base die that anchors the 12-high DRAM stack. For the first time, both SK hynix and Samsung have introduced a logic-enabled base die on 3nm and 4nm process technology to enable HBM4 product for efficient and faster product performance via base logic-driven memory management.
Both Samsung and SK hynix rely on TSMC for the production of their logic-enabled base die. However, it remains unclear whether Micron is using a logic base die, as the company lacks in-house capability to fabricate at 3nm.