NVIDIA seems to be having a tough time making up its mind about the final specs on the RTX 4070. Initially, it was rumored that the graphics card would come featuring a GDDR6 memory of 10 GB with a 160-bit bus interface. However, this was later updated to 12 GB with a 192-bit interface. As it turns out, both these options are on the table till now.
Famous leaker Kopite has been sharing specs on four different SKUs including the 340, 341, 336, and 337. Each of these variants comes with its own separate board design but it is important to note that the pairing of 340/341 and 336/337 denote the same SKU but only one is going to be used for the reference design. The other one goes off to a semi-custom design by the board partners.
As it turns out, the 340/341 is likely to be based on the full AD104 GPU (AD104-400) with a total of 7680 FP32 CUDA cores. This particular model will come with a GRRD6X memory of 12 GB which will be spread across a 192-bit interface and will have a speed of 21 Gbps, meaning a maximum bandwidth of 504 GB/s.
The 336/337 SKU on the other hand is based on a cut-down AD104 GPU which will give 56 streaming multiprocessors and a total of 7168 CUDA cores. Moreover, the memory size of this SKU will be only 10 GB and thus a smaller bus interface of 160-bit. Both these models are listed with 21 Gbps speed but this model will get a slightly lower bandwidth.
Regardless of which variant you will choose, the TBP specs have been reduced on both as compared to the past rumors. Kopite had previously shared that the total board power is going to be 300W but now it turns out that this has been reduced to 285W and 250W.
Review the two possible specs of RTX 4070:
PG141-SKU340/341, 7680FP32, 12G 21Gbps GDDR6X, total card power 285W, TSE <11000.
PG141-SKU336/337, 7168FP32, 10G 21Gbps GDDR6X, total card power 250W, TSE <10000.
I found that I had made a typo before, mistaken wrote 341 as 331. 😅— kopite7kimi (@kopite7kimi) August 29, 2022