Recently, NVIDIA ceased the shipment of high-performance GPUs to countries like China. The new U.S. restrictions aim to curb the utilization of powerful accelerators for specific tasks, particularly in machine learning and its potential use for military applications. This led to the export ban of cards such as H100, A100, and their 800 variants, originally designed to navigate previous restrictions. Notably, the updated rules shift the focus from interconnect communications between GPUs to compute power, impacting not only data-center GPUs but also gaming cards like RTX 4090.
In response, NVIDIA has today introduced a fresh series of data-center GPUs designed to comply with the new restrictions. The HGX H20 and L20 series are now officially available to Chinese customers, featuring reduced compute power compared to their predecessors. The H20 GPU boasts 96GB of HBM3 memory with a memory bandwidth of 4.0 TB/s. Interestingly, that’s higher bandwidth than the ‘global’ H100 with 3.6 TB/s.
On the other hand, the L20 GPU, based on an AD102 GPU, is equipped with 48GB of GDDR6 memory. These solutions intentionally limit compute capabilities to ensure GPUs do not exceed 4800 TOPS performance.
NVIDIA HGX H20, L20, L2 Specs, Source: ITHome
The HGX H20, a data-center GPU optimized for clusters of 8 GPUs, features NVLink connectivity and demands 400W of power. The L20 and L2 GPUs, operating on PCIe interface, offer memory options of 48GB and 24GB.
Reports indicate that NVIDIA has initiated chip sampling to partners, with orders set to commence from November 16th, coinciding with the full implementation of U.S. restrictions. NVIDIA anticipates the first shipment to take place in December.
