Earlier this month, the HBF memory standard was proposed. SK Hynix and Sandisk were directly involved in the development of the standard, the second of which has announced that it has a ready-made digital project that will allow it to start producing corresponding chips in the foreseeable future. Sandisk will reportedly provide the first physical samples of HBF next year, with mass production starting in 2028.
Image source: SanDisk
Sandisk representatives acknowledged this investor eventsThis happened yesterday. Based on their demonstration results, it is clear that HBF-type memory can provide the information transfer speed for HBM’s read operation characteristics, but at the same time increase the capacity by six to eight times. Compared to the classic layout using only HBM memory, adding HBF to the GPU computing configuration can produce the same number of tokens using half the number of GPUs. When running large language models, the combination of four GPUs with HBF memory is sufficient to provide performance similar to that historically achieved by eight GPUs with HBM memory.
Technically, this gives us hope to reduce the shortage of HBM and GPUs and reduce the cost of creating AI computing infrastructure. However, developer interest in the field is growing much faster than the emergence of more efficient computing technology. With the same capacity, HBF is cheaper than HBM, but has higher overall capacity, although it is not as good as HBM in terms of information transmission delay. The bandwidth of both memories is comparable. Most likely, in the future, both types of memory will coexist and combine, and HBF will not be able to replace HBM from the market. In the first case, the NAND wafers are combined into vertical stacks of 8 or 16, with a total volume of up to 512GB.
If you find an error, select it with your mouse and press CTRL+ENTER.
