NVIDIA launched its H100 “Hopper” data center GPU earlier this year. Boasting a 4nm node, and integrating up to 18K shaders, this monster GPU is the absolute cutting-edge accelerator out there. Like its predecessors, the focus has once again been matrix multiplication and mixed-precision compute which allows for a staggering throughput of up to 4,000 TFLOPs.
One peculiarity of the H100 is the unexpected use of TSMC’s 4nm node. The GeForce RTX 4080/4090 and the rest of Lovelace GPUs are also slated to leverage the same process which seemingly gives Team Green a minor advantage over AMD (The Radeon RX 7900 XT and 7800 XT will use both the 5nm and 6nm nodes to maximize performance while keeping the BOM down).
However, this is a common (deliberate?) misconception as both AMD and NVIDIA will be utilizing customized versions of the N5 node to better suit their needs. It’ll be hard to say which one is better, but you can be sure that the end result will be mostly the same on both ends.
This isn’t the first time TSMC has offered custom nodes to its clients. AMD has been using modified versions of TSMC’s N7 process over the last four years, with Zen 3 and Zen 3+ representing the true potential of fine-tuned process nodes. As already reported several times, this has allowed AMD’s Ryzen 6000 mobile processors to be substantially more power-efficient than Intel’s 12th Gen notebooks while also being faster at lower TDPs.