NVIDIA’s fourth Ada Lovelace graphics card is now available in retail. Featuring 5,888 FP32 cores and a bandwidth of 504 GB/s, it looks almost identical to its predecessor. The RTX 3070 packs 5,888 cores but pairs it with a wider 256-bit bus but slower GDDR6 memory for an external bandwidth of 448GB/s. On paper, the larger L2 cache and the higher memory bandwidth on the RTX 4070 are the only differences between the two 70-class graphics cards.
The above table from Quasarzone highlights how potent the Ada Lovelace microarchitecture is. Despite unchanged core counts, NVIDIA extracted up to 40% more performance from the RTX 4070 with just a node shrink (+higher clocks) and a larger LLC.
In real-life scenarios, the RTX 4070 offers roughly RTX 3080 level of performance, a GPU that has nearly 50% more cores (8,704) with a 50% higher bandwidth (760GB/s). These numbers go on to demonstrate the kind of generational uplift that NVIDIA was able to bring without changing the shader strength.