The American giant has presented its new memory HBM3 Gen2 memoryand has confirmed that it has already started shipping the first sample units, which means that it is only one step away from starting the mass manufacturing and distribution process, as long as everything goes according to plan.
Micron’s HBM3 Gen2 memory is available in a size of 24 GB in eight heights, offers a bandwidth greater than 1.2 TB/s and has a speed per pin of 9.2 Gb/s. These data represent an improvement of 50% compared to the current generation of HBM3 memory, and in terms of efficiency we are talking about a performance per watt that is 2.5 times better than previous generations.
Thanks to this new type of memory HBM3 Gen2 it will be possible reduce the training times of large language models based on artificial intelligencesuch as GPT-4, and more efficient infrastructures can be created to work with artificial intelligence and with different inference processes.
Micron has also confirmed the development of versions with a 36 GB capacity in 12 heights, although this is not yet available. It has also dropped in its roadmap the first details of its next-generation HBM memory, which will be available in 36 GB and 64 GB versions, and which will be able to reach a maximum bandwidth of 2 TB/s.
As we can read in the screenshot that shows the key specifications of Micron’s HBM3 Gen2 memory, it comes with error correction technology (ECC), and also has other state-of-the-art technologies that guarantee error-free operation and guarantee a long useful life. Among them we can highlight those for temporary or permanent repair of memory cells, and also automatic error checking.
As you may have imagined, this type of memory will be used in high-performance solutions aimed at the professional sector, such as next-generation graphics accelerators from both NVIDIA and AMD. For example, the NVIDIA H100, based on the Hopper architecture, uses HBM2e memory, while the Radeon Instinct MI300 uses HBM3 memory.
