With the RTX 30 series lineup, all four high-end SKUs were based on the same die, resulting in minuscule performance gains when going from the RTX 3080 to the 3080 Ti, and the 3090 to the 3090 Ti. For the first time, both the x80 (+ x80 Ti) and the Titan/x90 were powered by the same die, coming within single-digit performance figures of each other. Things will be returning to the convention with the upcoming RTX 40 series “Lovelace” graphics cards. The RTX 4080 and 4090 will be based on different dies, namely the AD103, and the AD102, respectively.
The GeForce RTX 4090 will come with up to 18,432 cores distributed across 144 SMs, 72 TPCs, and 12 GPCs, resulting in a single-precision throughput of nearly 90 TFLOPs. The GPU will be fed by a 24GB 24Gbps memory buffer via a 384-bit bus, supplemented by 96MB of L2 cache. It will have a peak power consumption of up to 600W.
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 GDDR6X | 16GB GDDR6 |
L2 Cache | 6MB | 6MB | 96MB | 64MB | 48MB |
Bus Width | 384-bit | 384-bit | 384-bit | 256-bit | 192-bit |
TGP | 250W | 350W | 600W? | 350W? | 250W? |
Launch | Sep 2018 | Sep 2020 | Aug-Sep 2022 | Q4 2022 | Q4 2022 |
The RTX 4080, on the other hand, will rely on the AD103 die, packing up to 10,752 cores, a moderate gain over the RTX 3080’s 8,704 FP32 cores. However, it’s important to note that this could refer to not just the FP32 ALUs but the INT32 ALUs as well. With Ampere, there were two separate data paths for INT and FP compute, with the former getting half as many units.
In terms of memory, we’re looking at 16GB GDDR6X memory paired with a 256-bit bus and supplemented by 64MB of L2 cache, much like the Radeon RX 6800 XT and the upcoming 7800 XT. The power draw will be similar to that of the RTX 4090, somewhere between 450-550W.
Finally, we have the RTX 4070 which will be yet another step below the RTX 4080. This mainstream GPU will leverage the AD104 die, racking in (up to) 7,680 shaders, a notable step up from the RTX 3070’s 5888 cores. These seven thousand something cores will be spread across 60 SMs and 5 GPCs and paired with 12GB of GDDR6 memory across a 192-bit bus. Once again, NVIDIA has cut down the bus width in favor of a larger L2 cache, 48MB in the case of the RTX 4070. The TGP is expected to be a rather digestible 300W vs 220W on the RTX 3070.
The RTX 4080 and 4090 are expected to be announced sometime in August or September, followed by the RTX 4070 in the holiday season. Expect a hefty price rise for the top-end SKUs.