Micron has announced that it has begun mass production of HBM3E, which will be used in Nvidia’s H200. This GPU is scheduled to start shipping in the second quarter of 2024. Micron states that this milestone positions it at the forefront of the industry, offering support for artificial intelligence (AI) solutions with industry-leading HBM3E performance and energy efficiency.
As early as July last year, Micron announced the launch of the industry’s first HBM3E with a bandwidth of over 1.2TB/s, pin speed of over 9.2GB/s, and a vertically stacked 24GB capacity using its 1-beta (1β) manufacturing process, which offers a 50% performance increase compared to the currently shipped HBM3 solutions. According to Micron, the per-watt performance of its HBM3E products is 2.5 times that of previous generations, setting new records for key metrics such as performance, capacity, and power efficiency in AI data centers. This can reduce training times for large language models like GPT-4 and offers superior total cost of ownership (TCO). Micron claims that its HBM3E’s power consumption is 30% lower than that of its competitors, SK Hynix and Samsung.
Micron has also prepared a 36GB capacity HBM3E with a 12-layer vertical stack that offers a 50% increase in capacity within the given stack height. As technology advances, Micron has doubled the silicon through-vias (TSVs) and increased the metal density by five times, thereby reducing thermal resistance. Combined with an energy-saving data path design, this has led to improved energy efficiency.
Next month at Nvidia’s GTC 2024, Micron will present the 36GB capacity HBM3E and share more product information.
Related:
