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 all NVIDIA GPU families in recent years, so we have little doubt that this leak is also accurate. We not only have the basic CUDA count with us but a lot of other details as well – so let’s dive into it.
The NVIDIA RTX 4070 GPU will rock 5888 CUDA cores for an estimated 30.7 FP32 TFLOPs, making it 76% faster than its big brother.
The RTX 4070 Ti and RTX 4070 are both based on the AD104 GPU and the 4070 will be the blocking variant for the RTX 4070 Ti GPUs. For our readers who aren’t familiar with binning: all GPU manufacturers actually produce the highest spec SKU but not all SM or CUDA cores are end lithographic working. So molds that work perfectly are sold as a higher end SKU and molds with fewer working parts get laser-scrambled or tape-scrambled and sold as a lower end product. This is also how AMD and Intel operate and it is part of the production science involved in semiconductor manufacturing.
Some RTX 4070 updates:
PG141-SKU336/337
AD104-250-A1
5888FP32
12GB 21GB GDDR6X
36m L2
250 watts– kopite7kimi (@kopite7kimi) December 9, 2022
According to Kopite7Kimi, the RTX 4070 will be a die cut AD104 with identifiers PG141-SKU336 and PG141-SKU337. The exact GPU variant name will be AD104-250-A1. The GPU will have exactly the same transistor count as the RTX 4070 Ti but only 5,888 CUDA cores across 46 SMs will be active. It will have the same VRAM as the 4070 Ti with 12GB of GDDR6X memory and the same bus width at 192-bit. The memory will also be clocked at the same 21Gbps which brings the bandwidth to a solid 504GB/s. This will be the first NVIDIA card to have a default TGP of 250W. The die size will be the same as the RTX 4070 Ti at 295mm². All of NVIDIA’s Ada Lovelace lineup is manufactured on TSMC’s 4nm process.
| Graphics card name | NVIDIA GeForce RTX 4090 | NVIDIA GeForce RTX 4080 | NVIDIA GeForce RTX 4070 Ti graphics card | NVIDIA GeForce RTX 4070 (estimated) |
|---|---|---|---|---|
| GPU name | Ada Lovelace AD102-300 | Ada Lovelace AD103-300 | Ada Lovelace AD104-400 | Ada Lovelace AD104-250 |
| process node | TSMC4N | TSMC4N | TSMC4N | TSMC4N |
| die size | 608 mm2 | 378.6 mm2 | 294.5 mm2 | 294.5 mm2 |
| transistors | 76 billion | 45.9 billion | 35.8 billion | 35.8 billion |
| CUDA cores | 16384 | 9728 | 7680 | 5888 |
| Strained/RT cores | 512/128 | 304/76 | 240/60 | 184/46 |
| base clock | 2230 MHz | 2210 MHz | 2310 MHz | 2310 MHz |
| clock increase | 2520 MHz | 2510 MHz | 2610 MHz | 2610 MHz |
| FP32 account | 83 TFLOPs | 49 TFLOPs | 40 TFLOPs | 30.7 TFLOPs |
| RTFLOPs | 191 TFLOPs | 113 TFLOPs | 82 TFLOPs | 62.9 TFLOPs |
| memory capacity | 24 GB GDDR6X | 16 GB GDDR6X | 12 GB GDDR6X | 2610 MHz |
| memory bus | 384 bits | 256 bits | 192 bits | 192 bits |
| memory speed | 21.0 Gbps | 23.0 Gbps | 21.0 Gbps | 21.0 Gbps |
| bandwidth | 1008 GB / s | 736 GB / s | 504 GB / s | 504 GB / s |
| TBP | 450 watts | 320 watts | 285 watts | 250 watts |
| price (current) | $1,599 | $1199 | $899 (TBC) | to be announced later on |
| launch (availability) | October 12, 2022 | November 16, 2022 | January 5, 2023 | to be announced later on |
Based on the core count specification mentioned by Kopite, we can now safely reverse engineer this card’s estimated FP32 performance. 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 from the same IHV and family, we can use the FP32 performance to directly compare the rasterization performance between the two GPUs and based on that, it looks like it will be about 76% faster than the RTX 4070 Ti. Of course, performance is just as important as pricing, and depending on how NVIDIA prices it, it may end up being a crowd favorite. Based on NVIDIA’s philosophy with pricing to date (and assuming no price cuts), the NVIDIA RTX 4070 could get you an MSRP of around $699 although this is pure speculation on my part.