NVIDIA’s flagship GeForce RTX 40 Series graphics card will be twice as fast as its BFGPU and current GeForce champion, the GeForce RTX 3090. This is according to rumors overheard by Twitter user greymon55, who noted that green team’s Ampere successors would leverage a 5-nanometer process and be released sometime between Q4 2022 and Q1 2023. While greymon55’s tweet doesn’t hold much weight on its own, notable leaker kopite7kimi has chimed in with a brief but corroborative comment that suggests NVIDIA will have no problem crushing the performance of its current flagship with its next cutting-edge architecture, Lovelace.
easy
— kopite7kimi (@kopite7kimi) July 6, 2021
| Lovelace AD102 | Ampere GA102 | |
|---|---|---|
| Process | Samsung 5 nm (?) | Samung 8 nm |
| Graphics Processing Clusters | 12 | 7 |
| Texture Processing Clusters | 72 | 42 |
| Streaming Multiprocessors | 144 | 84 |
| CUDA Cores | 18,432 | 10,752 |
| FP32 at 1.8 GHz | 66.4 TFLOPs | 38.7 TFLOPs |
| Memory Bus | 384-bit (?) | 384-bit |
| Memory Type | GDDR6X (?) | GDDR6X |
[…] a promising gain over the current Ampere range and lends further weight to the alleged “RTX 4080/4090” specs that were making news headlines back in December 2020. In terms of FP32 at 1.8 GHz, the Ampere GA102 chip offers 38.7 TFLOPs while the Lovelace AD102 reportedly computes at 66.4 TFLOPs. […] With a supposed 18,432 CUDA cores, compared to 10,752 for the GeForce RTX 3090, the top Lovelace-based card already holds a lot of performance promise.
Sources: greymon55, kopite7kimi, Notebookcheck