NVIDIA’s next-gen Lovelace graphics cards are rumored to feature considerably higher TGPs than their predecessors. We’re looking at power draws of up to 600W for the RTX 4090 and even more for the RTX 4090 Ti. This is despite the fact that these chips will be fabbed on TSMC’s 4nm N4 process which is easily one of the most efficient nodes on the planet. The excessive power consumption won’t be for nothing though, and the RTX 4090 (based on the AD102 die) will pack around 16K FP32 cores.
The AD102 die will get a haircut before going into the RTX 4090, dropping the core count from 18,432 to 16,384. This means that a few of the SMs, TPCs, and GPCs along with the L2 cache will also be axed. As for the clocks, Kopite7kimi states that we can expect core boost clocks well over 2.8GHz.
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 |
FP32 TF | 16.1 | 37.6 | ~100 TFLOPs? | ~50 TFLOPs | ~35 TFLOPs |
Memory | 11GB GDDR6 | 24GB GDDR6X | 24GB GDDR6X | 16GB GDDR6 | 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 |
We might get a GPU with 16,384 pulsating cores running at a whopping 3GHz and custom liquid-cooled models clocked even higher. The more accessible RTX 4080 featuring the AD103 die should pack around 10,000 cores and boosts exceeding 3GHz. The peak power draw should stay in the 400-450W range.
Clock speeds of 3GHz should allow the RTX 3090 to push through the 100 TFLOPs barrier albeit with significant heat generation and an unreasonable power draw. The RTX 4080 will hover a bit over (or under) 50 TFLOPs while the RTX 4070 should be rated over 30 TFLOPs. Keep in mind that these figures don’t mean much in terms of real world performance but are a reminder of how far GPUs have come.