NVIDIA already has new data-center GPUs for China
NVIDIA responded shortly to new restrictions on exporting highly effective GPUs from the U.S.
Not too long ago, NVIDIA ceased the cargo of high-performance GPUs to nations like China. The brand new U.S. restrictions purpose to curb the utilization of highly effective accelerators for particular duties, notably in machine studying and its potential use for navy purposes. This led to the export ban of playing cards similar to H100, A100, and their 800 variants, initially designed to navigate earlier restrictions. Notably, the up to date guidelines shift the main focus from interconnect communications between GPUs to compute energy, impacting not solely data-center GPUs but additionally gaming playing cards like RTX 4090.
In response, NVIDIA has at present launched a recent sequence of data-center GPUs designed to adjust to the brand new restrictions. The HGX H20 and L20 sequence at the moment are formally out there to Chinese language prospects, that includes diminished compute energy in comparison with their predecessors. The H20 GPU boasts 96GB of HBM3 reminiscence with a reminiscence bandwidth of 4.0 TB/s. Apparently, that’s increased bandwidth than the ‘international’ H100 with 3.6 TB/s.
Alternatively, the L20 GPU, based mostly on an AD102 GPU, is supplied with 48GB of GDDR6 reminiscence. These options deliberately restrict compute capabilities to make sure GPUs don’t exceed 4800 TOPS efficiency.
NVIDIA HGX H20, L20, L2 Specs, Supply: ITHome
The HGX H20, a data-center GPU optimized for clusters of 8 GPUs, options NVLink connectivity and calls for 400W of energy. The L20 and L2 GPUs, working on PCIe interface, provide reminiscence choices of 48GB and 24GB.
Experiences point out that NVIDIA has initiated chip sampling to companions, with orders set to start from November sixteenth, coinciding with the complete implementation of U.S. restrictions. NVIDIA anticipates the primary cargo to happen in December.
Supply: ComputerBase
