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

HBM2

HBM2 is short for High Bandwidth Memory 2. It is a type of random-access memory characterized by a specific architecture. Unlike conventional memory modules such as DDR4, HBM2 features individual memory chips stacked vertically on top of each other within a single package. This arrangement allows for the creation of an extremely wide data bus, which is the path along which data flows between the memory and the processor. As a result, HBM2 memory can transfer enormous amounts of data in a short amount of time. Furthermore, the memory chips are placed very close to the main computing chip, such as a graphics processor, on a common silicon substrate called an „interposer“. This proximity reduces the distance data must travel, leading to lower power consumption and faster communication. In AI accelerators and workstations, it prevents the powerful processor from having to wait for data, thereby maximizing its utilization during demanding computations, such as training neural networks. 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