NVIDIA launched its GeForce RTX 40 series graphics cards, codenamed Ada Lovelace, late last year. The GeForce RTX 4090, 4080, 4070 Ti, and 4070 are already out, with the 4060 Ti on the way. The Ada Lovelace family leverages TSMC’s N4 (custom 5nm) node consisting of five dies ranging from 144mm^2 (AD107) to 608mm^2 (AD102). The top-end AD102 GPU features 18,432 shaders across 144 SMs with 96MB of L2 cache.
I was wrong.
At the backend, you’re looking at 192 ROPs alongside 576 TMUs, and 2,304 SFUs. The AD102 die powers the GeForce RTX 4090 with a TBP of up to 450W. The former has a peak power limit of 800W and was slated to power the RTX Titan (before it was shelved).
The upcoming RTX 50 series GPUs, codenamed Blackwell, are planned for a late 2024 release. Like Ada, it will also be a monolithic design. According to Kopite7kimi, these GPUs will be fabbed on TSMC’s 3nm (N3) node. You can expect transistor counts of over a billion and densities of nearly 150 billion/mm². Blackwell should offer core clocks of over 3GHz and bus widths of up to 512 bits, resulting in memory bandwidths rivaling HBM memory.