After the failed acquisition of Arm, NVIDIA has decided to go ahead with its Grace CPUs in a bid to undercut its x86 rivals. When paired with the upcoming Hopper GPUs, these next-gen data center processors offer unparalleled levels of GPU-to-GPU and GPU-to-CPU bandwidth. More importantly, it allows the chipmaker to improve profit margins and strengthen its grip over the lucrative server/data center markets.
Last Tuesday, NVIDIA revealed its new graphics architecture “Hopper” and a slew of products based on the H100 GPU. Based on TSMC’s N4/N5P process node, this next-gen Tensor core accelerator features a whopping 80 billion transistors.
NVIDIA CEO Jensen Huang is optimistic about the company’s supply chain which presently leverages the Samsung-TSMC duology. Huang is interested in using Intel’s advanced process nodes in the future, namely the 20A and 18A (or the 2nm, and 1.8nm) technologies as well. However, any such relationship isn’t expected to mature anytime soon. At the earliest, NVIDIA and Intel might sign a foundry contract in 2024 for the 20A process node.