Try our cookies Alza.cz a. s., Company identification number 27082440, use cookies and other data to ensure the proper functioning of the website and, with your consent, also, among other things, to personalize advertising and the content of our websites. By clicking on the “I understand“ button, you agree to the use of cookies and the transfer of data regarding the behavior on the website for displaying targeted advertising on social networks and advertising networks on other websites.
I understand Detailed settings Reject everything
Alzak icon

HBM2e

HBM2e is a high-performance computer memory standard, short for 'High Bandwidth Memory 2 enhanced'. Unlike conventional memories, where chips are placed side-by-side on a board, HBM memory chips are stacked vertically. This arrangement enables the creation of a very wide communication path, known as a memory bus, capable of transferring large amounts of data simultaneously. Furthermore, the memory is placed on a common base, a silicon substrate called an 'interposer', in close proximity to the main computational chip, such as a Graphics Processing Unit (GPU). The short distance reduces data transfer latency and power consumption. In the field of artificial intelligence, high data throughput is crucial for efficiently supplying data to processing units, preventing their idle time and allowing their full performance to be utilized during the training and operation of complex models. The letter 'e' in the name denotes an enhanced version of the HBM2 standard, offering higher transfer rates. To understand how it differs from conventional gaming graphics memory, a comparison with GDDR (Graphics Double Data Rate), such as GDDR6 or GDDR7, is helpful. GDDR consists of individual chips placed on the board around the graphics core and relies on very high clock speeds over a relatively narrow bus (typically 32 bits per chip). Memories in the HBM family take the opposite approach. They run at lower clock speeds but, thanks to vertically stacked chips and an extremely wide bus (1024 bits or more per stack), they transfer more data at once. The result is significantly higher total data throughput for HBM and a better performance-per-watt ratio, which is crucial for data centres and AI training. GDDR, by contrast, is cheaper, easier to manufacture and easier to scale in capacity, which is why it dominates in gaming and consumer graphics cards.

Offer of AI Accelerators Offer of AI Workstations
Print
P-DC1-WEB23