The upcoming release of NVIDIA's RTX 40 series of graphics cards has been heavily speculated about for quite some time. Now, kopite7kimi provides information on NVIDIA's newest graphics cards, the RTX 4070, RTX 4080, and RTX 4090.
AD102-300-A1 GPU will be used in the GeForce RTX 4090, which would get 16,384 shader cores and 24GB of GDDR6X memory running at 21Gbps on a 384-bit bus. There will be a boost from its base frequency of 2235 MHz on the GPU to a maximum of 2750 MHz. Note that the AD102 chip actually would have 18,432 FP32 cores when fully activated. And as far as we can speculate, the GeForce RTX 4090 Ti will be the first to use them, providing performance of up to 100 TFLOPs. On the negative side, and this is glaringly clear, the TGP is probably limited to 600W.
The GeForce RTX 4080 will include an AD103-300-A1 GPU with 10,240 FP32 cores and 16GB of 21Gbps GDDR6X memory on a 256-bit bus. The core of the GPU will be clocked at about ~2.5GHz, with a possible boost surpassing 3GHz. The RTX 4080 will feature a reasonable TGP of 420W, which will be at least 50 percent more than the RTX 3090. It should cost around $999.
The GeForce RTX 4070, equipped with the AD104 GPU, will contain 7168 shaders and 10 GB of GDDR6 memory operating at 18 Gbps on a 160-bit bus. The GPU will be clocked between 2.5 and 3 GHz, while the TGP will be 300W. The RTX 4070 will probably cost $499, the same as the RTX 3070.
The L2 cache sizes of the RTX 4090, 4080, and 4070 are 96 MB, 64 MB, and 48 MB, respectively. Beginning with the RTX 4080 and 4090, these 4nm GPUs are planned to launch in September/October.
GPU | TU102 | GA102 | AD102 | AD103 | AD104 |
---|---|---|---|---|---|
Architecture | Turing | Ampere | Ada Lovelace | Ada Lovelace | Ada Lovelace |
Fabrication node | TSMC 12nm | Samsung 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 | 4608 | 10752 | 18432 | 10752 | 7680 |
TP | 16.1 | 37.6 | ~90 TFLOPs | ~50 TFLOPs | ~35 TFLOPs |
Memory | 11GB GDDR6 | 24GB GDDR6X | 24GB GDDR6X | 16GB GDDR6 | 16GB GDDR6 |
L2 Cache | 6MB | 6MB | 96MB | 64MB | 48MB |
Bus | 384-bit | 384-bit | 384-bit | 256-bit | 192-bit |
TGP | 250W | 350W | 600W? | 350W? | 250W? |
Launch | Sep 2018 | Sep 2020 | August-Sep | Q4 2022 | Q4 2022 |
Rumor: GeForce RTX 4090, RTX 4080 and RTX 4070 > 16384 cores, 24GB GDDR6X, 2.7 GHz, 450W TGP