NVIDIA RTX 4070 Specifications Leak Out: 250W Affair With 5888 CUDA Cores, 12GB GDDR6X Memory And More

One of the most accurate and influential NVIDIA leakers, @kopite7kimi has leaked the specifications of NVIDIA's upcoming GeForce RTX 4070 graphics card. Kopite was the first to leak the existence and specifications of pretty much all NVIDIA GPU families in recent years so we have very little doubt that this leak is accurate as well. We not only have the CUDA core counts with us but a lot of other details as well - so let's dig in.

The RTX 4070 Ti and RTX 4070 are based on the AD104 GPU and the 4070 will be the binned variant of the RTX 4070 Ti GPUs. For our readers who are not familiar with binning: all GPU manufacturers actually produce the higher spec-ed SKU but not every SM or CUDA core is functional by the end of the lithographic process. So the perfectly working dies are sold as the higher end SKU and the dies which have fewer functioning parts get the dysfunctional parts lasered off or disabled with straps and sold as a lower-segment product. This is also how AMD and Intel operate and is part of the yield sciences involved in semiconductor manufacturing.

Some updates of RTX 4070: PG141-SKU336/337AD104-250-A15888FP3212G 21Gbps GDDR6X36M L2250W

— kopite7kimi (@kopite7kimi) December 9, 2022

According to Kopite7Kimi, the RTX 4070 will be a cut-down AD104 die with ID of PG141-SKU336 and PG141-SKU337. The exact GPU variant name is going to be AD104-250-A1. The GPU will have the exact same transistor count as the RTX 4070 Ti but only 5888 CUDA Cores will be activated across 46 SMs. It will feature the exact same VRAM as the 4070 Ti at 12GBs worth of GDDR6X memory and the same bus width at 192-bit. The memory will also be clocked at the same rate of 21 Gbps bringing the bandwidth to a solid 504 GB/s. This is going to be the first card from NVIDIA that is going to have a default TGP of 250W. The die size is going to be the same as the RTX 4070 Ti at 295mm². All of NVIDIA's  Ada Lovelace lineup is manufactured on TSMC's 4nm process.

Based on the core count specifications Kopite mentioned, we can now safely reverse engineer the estimated FP32 performance of this card. The card should have a nominal FP32 performance of 30.7 TFLOPs (5888 CUDA Cores * 2 * 2.610 GHz = 30.7 TFLOPs). Since this is a card of the exact same IHV and family, we can use the FP32 performance to directly compare rasterization performance between GPUs and based on this, it looks like it is going to be about 76% as fast as an RTX 4070 Ti. Of course, the performance is only as important as the pricing and depending on how NVIDIA prices this, it could end up being a crowd favorite. Based on NVIDIA's philosophy with pricing so far (and without assuming any price cuts) the NVIDIA RTX 4070 could have an MSRP of around $699 although this is pure speculation on my part.