In the latest tweet, @ kopite7kimi disclosed more details of NVIDIA's upcoming geforce RTX 4090 / 4070 graphics card based on ADA lovalace GPU architecture Compared with geforce RTX 30 series graphics card based on ampere GPU, its specifications, performance, power consumption and other parameters are exaggerated.
（via WCCFTech）
For example, it is rumored that the flagship RTX 4090 is equipped with 24GB @ 21gbps gddr6x video memory, and the power consumption has reached an amazing 600W. In addition, the RTX 4070 is equipped with 12gb @ 18gbps gddr6 video memory, and its power consumption is as high as 300W.
Previously @ kopite7kimi also disclosed the specifications of other graphics cards, such as RTX 4090 Ti / 4080. The RTX 4090 will be based on the optimized ad102 GPU core, and the RTX 4070 will use the optimized AD104 GPU core.
It is rumored that the RTX 4090 also uses the same ad102 GPU as the 4090 Ti, but Lao Huang, with exquisite knife technique, left the full blood version to the latter.
Compared with ga102 GPU, ad102 GPU has 144 groups of SM units (+ 71%) / 18432 CUDA cores, and ADA Lovelace GPU also increases L2 cache.
The flagship ad102 GPU has 96MB L2 cache, which has soared to 16 times that of the previous generation ga102 GPU (with 6MB L2). In addition, the bandwidth of 384 bit @ 21gbps 24GB gddr6x video memory also exceeds 1.0 TB / s.
In terms of power consumption, given that the TBP data of the public version of RTX 3090 Ti has exceeded 450W / that of the non-public version has reached 516w, I am not surprised by the disclosure of the super 600W TBP of ADA Lovelace flagship (RTX 4090).
In order to better suppress the heat output of the chip (TBP is 33% higher than RTX 3090 Ti), AIB manufacturers may introduce more hybrid heat dissipation design (instead of relying on the traditional three slot multi fan radiator).
As for the RTX 4070 with AD104 GPU (said to be pg141-sku341 PCB), and the graphics card has 60 groups of SM units / 7680 CUDA cores / 48MB L2 cache.
In terms of video memory, the card may adopt 192 bit @ 16-18gbps gddr6 specification, with a bandwidth of about 400 GB / s, supplemented by up to 300W TGP power consumption.
Since the TDP of the public RTX 3070 Ti has reached 290w, the power consumption increase of only 10W is negligible, but its performance is equivalent to that of the 350W RTX 3080 ti.
If NVIDIA can adopt a more competitive pricing strategy for RTX 40 Series graphics cards, the RTX 4070 of $500-600 (3080 Ti of $1199) will undoubtedly be regarded as a huge upgrade.