Most impressive, of course, are the specs of the flagship GeForce RTX 3090. This graphics card is built around an Ampere GPU with 10,496 CUDA cores. At the same time, the previous flagship Titan RTX based on Turing offers only 4608 CUDA cores, and in all early leaks of the characteristics of gaming Ampere for the GeForce RTX 3090, 5248 CUDA cores were indicated.

If we turn to the official specifications of other cards, it turns out that the GeForce RTX 3080 is built on a chip with 8704 CUDA cores, while the GeForce RTX 2080 Super chip has only 3072 CUDA cores. Finally, the youngest of the presented new products – GeForce RTX 3070 – according to NVIDIA has 5888 CUDA cores, which is also more than any previous generation card.

Unfortunately, the number of RT cores for ray tracing and tensor cores for AI tasks is not specified, but there are probably a lot of them here. It is only noted that the gaming Ampere uses second-generation RT cores and third-generation tensor cores. The reference Boost frequency for the older GPU reaches 1700 MHz, for the middle one – 1710 MHz, and for the younger one – 1730 MHz.

In general, it can be noted that the increase in the number of CUDA cores during the transition from Turing to Ampere was about 2.5 times. And this is one of the reasons for the significant increase in performance that new items promise.

Speaking of CUDA cores, it should also be noted that the flagship server GPU of the Ampere generation – NVIDIA A100 – has significantly fewer shaders. This is explained by the different structure of gaming GPUs and versions of chips for data centers. In gaming GPUs, NVIDIA has purposefully increased the number of FP32 cores as today’s ray-traced gaming workload requires additional computing resources to be spent on noise reduction and reflection shading.