NVIDIA’s AD104 die is about to start testing at NVIDIA’s labs soon. This indicates that the final specifications of the next-gen GeForce RTX 4070 will be ironed out and finalized. The AD104 die will consist of 7,680 FP32 cores across 60 SMs (Streaming Multiprocessor) and 30 TPCs (Texture Processing Cluster). These will in turn be spread across 5 GPCs (Graphics Processing Cluster).
| GPU | TU102 | GA102 | AD102 | AD103 | AD104 |
|---|---|---|---|---|---|
| Arch | Turing | Ampere | Ada Lovelace | Ada Lovelace | Ada Lovelace |
| Process | TSMC 12nm | Sam 8nm LPP | TSMC 5nm | TSMC 5nm | TSMC 5nm |
| GPC | 6 | 7 | 12 | 7 | 5 |
| TPC | 36 | 42 | 72 | 42 | 30 |
| SMs | 72 | 84 | 144 | 84 | 60 |
| Shaders | 4,608 | 10,752 | 18,432 | 10,752 | 7,680 |
| TP | 16.1 | 37.6 | ~90 TFLOPs? | ~50 TFLOPs | ~35 TFLOPs |
| Memory | 11GB GDDR6 | 24GB GDDR6X | 24GB GDDR6X | 16GB GDDR6 | 12GB/16GB GDDR6 |
| L2 Cache | 6MB | 6MB | 96MB | 64MB | 48MB |
| Bus Width | 384-bit | 384-bit | 384-bit | 256-bit | 192-bit/256-bit |
| TGP | 250W | 350W | 600W? | 350W? | 250W? |
| Launch | Sep 2018 | Sep 2020 | Aug-Sep 2022 | Q4 2022 | Q4 2022 |
The AD104 die will be paired with a 192-bit bus which means we’re looking at a 12GB VRAM buffer along with 48MB of L2 cache. There’s a chance we’ll get a 256-bit bus as well, and in that case, expect 16GB of GDDR6 memory.
From the tape-out schedule, it would seem that NVIDIA plans on launching the GeForce RTX 4080 and 4090 in August, followed by the RTX 4070 in October/November. This makes sense as it’ll be more profitable to sell the pricey 5nm wafers to enthusiasts at higher profit margins first, and release the budget offerings later on as yields gradually stabilize.
