News

Nvidia H100 infrastructure noted for its use of older HBM2e and HBM3 memory standards

Saturday, March 28, 2026 at 09:10 AM

The tweet discusses older GPU generations, specifically referencing the H100 and its use of HBM2e and HBM3 memory compared to the newer H200. It notes the age of these components in the context of current infrastructure.

Context

As of March 28, 2026, market attention is shifting back to the legacy hardware powering established AI clusters. While Nvidia has moved into the Blackwell and Rubin eras, the H100 remains a foundational asset. Recent analysis highlights that the H100 SXM primarily utilizes HBM3 memory, while some PCIe variants and older infrastructure still rely on HBM2e. This distinguishes the H100 from the H200, which was the first to adopt the significantly faster HBM3e standard to address memory-bandwidth-bound workloads. This technical distinction is critical for infrastructure efficiency. The H100 delivers approximately 3.35 TB/s of bandwidth, compared to the 4.8 TB/s found in the H200. For investors, the continued use of HBM2e/3 in H100 systems reflects the long tail of the Hopper architecture's lifecycle and the supply chain's transition toward HBM4, which is expected to enter mass production by SK Hynix and Samsung later this year.

Related Companies

Nvidia
Nvidia
NVDA
US