The Benchmark
Over at Geekbench, someone tested the RTX 4090 before launch. The 4090 is paired along with AMD’s kingpin Ryzen 9 7950X on the ASRock X670E Taichi Carrara motherboard and DDR5–6000 memory. The RTX 4090 scores 417713 points in the CUDA test which is much higher than its predecessor. Of course, this is not an extremely common way to benchmark the performance since most GPUs are tested in 3DMark Time Spy Extreme. However, to put this resultl into perspective, we compiled a detailed CUDA chart from Pascal to Lovelace. As shown below, the uplift from the GTX 1080 Ti to the RTX 2080 Ti in terms of CUDA performance was a massive 183.45%. This result was achieved with the core count increasing by just 21.43%. For a more fair comparison, under the ‘RTX’ branding from the 2080 Ti to the 3090 Ti we see a CUDA core uplift of a whopping 147.06%. This time, the CUDA score saw a 71% increment from generation to generation. (The percentage in the grey bar represents the % Cuda Score increase as compared to the previous generation. The Orange bar shows the generation on generation Cuda Core count increase). As for Lovelace, the CUDA core count is increased by a modest 52.38%. The CUDA score, however, sees a 1:1 uplift at 51.57%. From Pascal -> Turing, the increase is explained by the architectural redesign. The following leap (Turing -> Ampere) sees a 71.41% increment but that needs almost 2x the CUDA core increase.
Conclusion
All things considered, Ada Lovelace seems to be much more different than Ampere than we initially thought. With Ampere, for a 70% CUDA performance increase, we needed almost 140% more cores. However, Lovelace offers more a linear increase. (X-axis represents the Cuda Core count | Y-axis represents the Cuda Core score)
