The RTX40 series of gaming notebooks will be on the market in a few days, and the merchant chose a special day, February 22 at 22:00, which is really thoughtful. In order to be targeted when choosing friends who use game laptops as AI, I specifically checked the CUDA cores and frequencies of the RTX40 series, and calculated FP32 TFLOPS for easy comparison.
Laptop GPU:
RTX40 Series Notebook GPUs:
RTX30 Series Notebook GPUs:
The RTX30 series of gaming notebooks still occupy a relatively large market. It is estimated that with the launch of the RTX40 series of gaming notebooks, there may be some discounts on the price, and the cost performance will be improved. You can also pay attention to it.
Comparison between RTX40 series and RTX30 series gaming notebooks:
Tests given by NVIDIA show that the performance of RTX40 series notebook GPUs has been greatly improved compared with the previous generation:
In addition to the increase in acceleration frequency, it also benefits from the upgrade of the architecture:
Comparison of computing power between RTX40 series and RTX30 series gaming laptops:
Desktop graphics card:
RTX40 Series Desktop GPUs:
RTX30 Series Desktop GPUs:
Comparison of RTX40 series and RTX30 series desktop graphics cards:
summary:
Comparing the RTX30 and RTX40 series gaming GPUs, it is not difficult to find that the TFLOPS of the slightly lower-end RTX4060 and RTX3060 is not much different, and the bit width of the RTX4060 is lower than that of the RTX3060. It is estimated that Huang’s knife is aimed at 10~20 TFLOPS Yes, it is almost enough. If it is higher, the heat dissipation and cost performance will not be easy to control; the FP32 of the high-end RTX4090 gaming notebook has reached nearly 40 TFLOPS; at present, it cannot be higher, and the heat dissipation will not hold if it is estimated to be higher. Living.
The FP32 of the RTX4090 desktop graphics card has reached 82.6 TFLOPS, and the RTX4070Ti has also reached 40.1 TFLOPS, surpassing the flagship RTX3090Ti of the previous generation.
ChatGPT, which is currently popular, has a pre-trained large model with 175 billion parameters at the bottom layer, and at least tens of thousands of Nvidia GPU A100 (19.5 TFLOPS) are needed to support its computing power infrastructure. The cost of one model training exceeds 12 million US dollars; the total computing power in the training phase The consumption is about 3640PF-days (that is, the efficiency of 1 PetaFLOP/s runs for 3640 days). If you use 121 latest RTX4090 desktop graphics cards and run non-stop, you will have to run for a whole year!
Lao Xu, 2023/2/17