a summary: The AI Accessor race leads to fast innovation in high -frequency memory techniques. At the GTC event for this year, the Samsung, SK Hynix and Micron giants inspected HBM4 and HBM4E solutions from the next generation.
While graphics processing units in the data center are transferred to HBM3E, the memory road maps that were detected in NVIDIA GTC show that HBM4 will be the next big step. The computer attended the event and pointed out that this new criterion Enabling Some of the dangerous density and display of the frequency range on HBM3.
SK Hynix offered the first HBM4 48GB Mix consisting of 16 layers of 3GB chips at 8 GB per second. Likewise, Samsung and Micro had highly high HBM4 offers, where Samsung claims that the speeds would eventually reach 9.2 GB per second in this generation. We must expect to become more than 12 GB chimneys of more than 12 GB to launch HBM4 products in 2026.
However, memory makers already look beyond HBM4 at HBM4E and amazing capacity points. The Samsung 32 GB Road Map requires each dramatic layer, allowing 48 GB to 64 GB per staple with data rates between 9.2-10 GB per second. SK Hynix alluded to 20 or more chimneys, allowing capabilities of up to 64 GB using their 3GB chips on HBM4E.
This high density is crucial for the processes of upcoming Rubin graphics processing units in NVIDIA that aims to train artificial intelligence. The company has unveiled the use of Rubin Ultra 16 components of HBM4E to get 1 tPs of per unit of graphics processing when it arrives in 2027. NVIDIA claims that with four numbers per package and a 4.6PB/S system, Rubin Ultra can be able to memory in NVL576.
While these numbers are impressive, they come at a high -teeth price. Videocardz notes that consumer graphics cards seem unlikely Adopt HBM variables any time soon.
The HBM4 and HBM4E generation represents an important bridge to enable the scaling of artificial intelligence. If memory makers can achieve their aggressive density and frequency domain maps during the next few years, this will be widely enhanced by the AI’s database. Nafidia and others depend on it.
Image credit: Computer base