According to one rumor, Nvidia’s next-gen graphics cards will offer up to 18,432 cores, almost twice as many as the RTX 3090.
The current generation of graphics cards in the form of AMD’s Radeon RX 6000 and Nvidia’s RTX 3000 is still hardly available, but the first rumors about the specifications of the next generation are already doing the rounds on the web.
Specifically, rough key data of a GPU based on Nvidia’s Ada-Lovelace architecture has surfaced. The basic design is supposed to be much more opulent than that of the current generation and thus offer more space for the arithmetic units.
Nvidia traditionally names its architectures after famous mathematicians. Ampere, for example, is named after André-Marie Ampère, and Turing after Alan Mathison Turing. Ada Lovelace was also a British mathematician who is also considered by some historians to be the first real programmer. The programming language Ada is also named after her.
The latest information comes from the leaker kopite7kimi ( @ kopite7kimi ), known in hardware circles , who, for example, predicted the specifications of Nvidia’s current amp architecture very accurately in mid-2019.
This is how the AD102 chip would look compared to its predecessors:
|Lovelace (RTX 4000) *||Chip: AD102||Manufacturing: 5nm||Cores: 18,432||TFLOPS: approx. 66||Release: 2022|
|Ampere (RTX 3000)||Chip: GA102||Manufacturing: 8nm||Cores: 10,752||TFLOPS: approx. 37||Release: 2020|
|Turing (GTX 1600 / RTX 2000)||Chip: TU102||Manufacturing: 12nm||Cores: 4,608||TFLOPS: approx. 16||Release: 2018|
|Pascal (GTX 1000)||Chip: GP102||Manufacturing: 16nm||Cores: 3,840||TFLOPS: approx. 12||Release: 2016|
If the specifications are fairly accurate, we will see a massive increase in core count and theoretical computing power with Nvidia Ada Lovelace. The latter in turn is based to a not insignificant degree on the clock rate, for which 3DCenter assumes a very conservative 1.8 GHz.
What is actually going on with Nvidia Hopper?
The first rumors about the Hopper architecture emerged more than a year ago. It was supposed to make use of the so-called MCM (multi-chip module) design, also called chiplet design, for the first time.
So, we might see Nvidia Hopper a bit later and get an intermediate step with Lovelace. It was similar with Ampere and Turing.
Rumors should be treated with caution: Nevertheless, all this is mere speculation for the time being. For a long time, the experts assumed that the chips for RTX 3000 would mainly be manufactured by Samsung in a 7-nanometer process; in the end, it became an 8-nanometer process from Samsung. Furthermore, hardly anyone had a quasi-doubling of the CUDA cores on their radar.