Besides, the RTX 4090 is very expensive
Nvidia decided to once again show how much all these modern NPUs in processors are inferior to the power of GeForce video cards in AI tasks.
For the demonstration, the company took the GeForce RTX 4090, which, on the one hand, demonstrates maximum capabilities, but on the other hand, makes the comparison not very correct, since the RTX 4090 costs more than $1,500.
Be that as it may, Nvidia’s flagship truly provides a completely different level of performance than NPUs in processors. In raw numbers we are talking about 1321 TOPS for a video card versus 10-20 TOPS in already released APUs and about 50 TOPS in those preparing for release.
In reality, as you can see in the diagrams, performance also differs significantly. In particular, Nvidia is helped here by its own TensorRT-LLM library, which became available for Windows at the end of last year.
Without this library, in the Jan.ai platform test, the video card outperforms the Ryzen 9 8945HS by nine times, and with it the advantage grows to 15 times.
By the way, if you look at a simpler video card — RTX 4070 Laptop — the difference is reduced to 4-5 times. And here it will be interesting to look at the new Ryzen AI 300 and Intel Lunar Lake APUs. It is quite possible that they will be quite close to the video card, but this is far from the most affordable mobile card from Nvidia.