NVIDIA’s RTX 4070 is going to be an absolute price-performance king. Based on the AD104 die, this upper midrange GPU will sport 7,168 FP32 cores paired with 10GB of GDDR6 memory across a 160-bit bus. The core will be clocked at a staggering 3GHz or above with a TGP of 300W, up from 220W on the RTX 3070. Compared to the RTX 4080 and 4090, it’ll get a much lower bandwidth but this shouldn’t be an issue at 1080p and 1440p.
|GPU||GA102||AD102||RTX 4090||AD103||RTX 4080||AD104||RTX 4070|
|Arch||Ampere||Ada Lovelace||Ada Lovelace||Ada Lovelace|
|Process||Sam 8nm LPP||TSMC 5nm||TSMC 5nm||TSMC 5nm|
|TP||37.6||~100 TFLOPs?||83 TFLOPs||~50 TFLOPs||47 TFLOPs?||~35 TFLOPs||35 TFLOPs?|
|Memory||24GB GDDR6X||48GB GDDR6X||24GB GDDR6X||16GB GDDR6X||12GB GDDR6X|
|Launch||Sep 2020||Sept 22?||Sept 22?||Q1 2023?|
According to Kimi, the AD104 die powering the RTX 4070 will roughly offer the same level of performance as the GA102 which was the chip behind the RTX 3080, 3080 Ti, 3090, and the 3090 Ti. Considering that the RTX 3090 packs 10,496 cores and a boost clock of around 1.6-1.7GHz, this doesn’t sound far-fetched. The increased core count of 7,168, boost clocks of over 3GHz, 48MB L2 cache, and other architectural and process (5nm) improvements should easily propel the RTX 4070 ahead of the RTX 4090, most notably in ray-tracing workloads.
Of course, despite the superior process and architecture, the power consumption will also increase by around 50% compared to 300W (just 50W less than the RTX 3090) which means you’ll need a 700-750W power supply for optimal operation.