A800/H800 models now actively promoted for other regions
NVIDIA not letting unsold H800 go to waste.
The recently imposed export restrictions have had an impact on companies like NVIDIA that provide high-performance accelerators for machine learning. These restrictions prevent the shipment of powerful chips to specific countries, including China. The new restrictions which were announced on October 16, already prevent NVIDIA from shipping a number of GPUs, including HPC and even gaming cards, such as RTX 4090.
Initially, a list of affected models included A100/H100 and even specialized versions of A800/H800 GPUs created by NVIDIA to overcome previous export limitations. Unlike the previous restrictions from 2022, the new sanctions focus on limiting compute performance rather than interconnect speed. Consequently, as of last week, NVIDIA is unable to ship the new A800/H800 models to China, leaving the company with a stock of cards that its primary market can no longer access.
In response, NVIDIA has redirected its H800 GPUs toward western markets, evident through new pages on the official NVIDIA website that promote these GPUs. Various companies that distribute NVIDIA’s HPC and workstation cards, including PNY, have begun offering these cards in North America, Latin America, Europe, Africa, and India. PNY is not the sole provider of these cards, as Japan-based companies like ASK Corporation and ELSA have also started offering these 800 models, as reported by CRN.
It is interesting that the way the A800 40GB card is promoted has changed. Previously positioned as a capable data-center GPU, it is now marketed as the ‘ultimate workstation development platform for AI’. The A800 GPU closely resembles the A100 version, with both cards featuring 6912 CUDA cores, 40GB of HBM2 memory, and a 240W TDP. The primary distinction lies in the reduced NVLink bandwidth, dropping from 600 to 400 GB/s. However, given the new constraints on compute power, this adjustment becomes a significant aspect.
This workstation-oriented version, termed “Active” by NVIDIA, resembles the design and cooling solutions of typical RTX models, with the exception of one missing feature. The A100 40GB Active lacks display connectors, limiting its usage to a compute coprocessor, necessitating the presence of a dedicated graphics card in the system.
The appearance of the A800 GPU on the global market will only allow NVIDIA to sell the remaining stock quicker. The A800 40GB Active model sales will start in the next few weeks. The US Distribution Executive cited by CRN believes that these cards will sell out fairly quickly.