Nvidia has placed an urgent order with TSMC to produce artificial intelligence (AI) and high-performance computing (HPC) GPUs for its Chinese customers with TSMC, according to a report Laoyaoba.com (via TrendForce). The orders reportedly fall under the category of 'Super Hot Run' and are set to be fulfilled in Q1 2024.
The report comes from an unofficial source and should be treated appropriately. Meanwhile, the key fact of the report is that Nvidia is proceeding with a line-up of AI and HPC GPUs designed specifically for China and to meet U.S. export regulations. Another aspect worth mentioning is that this line-up is now set to be introduced in Q1 2024 rather than December 2023. Perhaps, because Nvidia needs to ensure that the U.S. government is not restricting this product family.
What exactly TSMC is set to produce for Nvidia is something that remains to be seen. Nvidia's rumored lineup of AI and HPC GPUs aimed at Chinese market includes the HGX H20 GPU (SXM card, GH100 silicon, 96 GB HBM3 memory), L20 (PCIe card, AD102 silicon, 48 GB GDDR6 memory), and L2 (PCIe card, AD104 silicon, 24 GB GDDR6 memory) and all of them use existing silicon. To cut down their performance, Nvidia (other chip designers use similar methods) either changes microcode of the GPU, or blows certain fuses within the GPU (and alters microcode appropriately), it does not have to alter GPU design and produce new GPUs with alternative configurations.
That said, it is unclear what exactly TSMC is said to produce for Nvidia in addition to what it already makes for the company. Perhaps, now that Nvidia's H20, L20, and L2 GPUs offer lower performance than their counterparts aimed at the U.S. and Europe, Nvidia's China customers need more of China-specific GPUs to achieve desired performance levels. This may be why Nvidia had to place an urgent 'Super Hot Run' order with TSMC. Or maybe it somewhat altered the process of making cut down GPUs to speed up their production and/or lower costs (though the latter is very unlikely, given the costs associated with a "Super Hot Run" of new silicon). Also, Nvidia has billions of dollars of pre-purchased TSMC capacity, so, it is unclear whether it needs to place an urgent need at all.
Keep in mind that we are dealing with an unofficial piece of information and this one should be taken with a huge grain of salt.
