NVIDIA has reportedly cut the power consumption of its high-end Lovelace graphics cards. The RTX 4080 which was originally planned to be a 420W SKU has been scaled down to 320W while the TGP of the RTX 4070 has been reduced from 300W to 285W. With this final adjustment, the RTX 4080 is set to have the same power consumption as the RTX 3080 while the RTX 4070 will top out 65W north of the RTX 3070.
According to @kopite7kimi, the end performance of the two GPUs hasn’t dropped by much, and the use of superior boards (originally meant to handle up to 420W) will leave plenty of room for overclocking for enthusiasts and gamers alike.
| GPU | GA102 | AD102 | RTX 4090 | AD103 | RTX 4080 | RTX 4070 Ti (AD104) | RTX 4070 |
|---|---|---|---|---|---|---|---|
| Arch | Ampere | Ada Lovelace | Ada Lovelace | Ada Lovelace | |||
| Process | Sam 8nm LPP | TSMC 5nm | TSMC 5nm | TSMC 5nm | |||
| GPC | 7 | 12 | 11 | 7 | 7 | 5 | 5 |
| TPC | 42 | 72 | 64 | 42 | 40 | 30 | 30 |
| SMs | 84 | 144 | 128 | 84 | 80 | 60 | 60 |
| Shaders | 10,752 | 18,432 | 16,384 | 10,752 | 9,728 | 7,680 | 7,680 |
| TP | 37.6 | ~100 TFLOPs? | 83 TFLOPs | ~50 TFLOPs | 47 TFLOPs? | ~35 TFLOPs | 35 TFLOPs? |
| Memory | 24GB GDDR6X | 48GB GDDR6X | 24GB GDDR6X | 16GB GDDR6X | 12GB GDDR6X | ||
| L2 Cache | 6MB | 96MB | 72MB | 64MB | 48MB | ||
| Bus Width | 384-bit | 384-bit | 256-bit | 192-bit | |||
| TGP | 350W | 600W | 450W | 450W | 285-340W | 300W | 285W |
| Launch | Sep 2020 | Sept 22? | Sept 22? | Q1 2023? |
The RTX 4080 will leverage the AD103 die featuring 9,728 FP32 cores alongside 16GB of GDDR6X memory across a 256-bit bus. NVIDIA has disabled 1,024 FP32 cores or 8 SMs on the AD103 die, leaving room for an RTX 4080 Ti on the same possibly with a wider bus. It was originally supposed to feature a boost clock of over 3GHz but with this TGP cut, we’re likely looking at a 2.7-2.9GHz boost (AIB models approaching 3GHz). The 64MB of L2 cache will have a crucial impact on gaming performance.
The RTX 4070 will feature the fully enabled AD104 die: 7,680 shaders across 60 SMs paired with 12GB of GDDR6X memory across a 160-bit bus. Even this midrange Lovelace GPU will come with 48MB of L2 cache. With a TGP of 285W, it’s 65W more demanding on the power supply than its predecessor. At the moment, it’s unclear whether the 4070 will be rolled out with its higher-end brethren in September or delayed to the first half of 2023.
The RTX 4090 is the one GPU whose specifications are more or less confirmed. It’ll feature 16,384 cores alongside 24GB of GDDR6X memory across a 384-bit bus and 72MB of cache (down from the total of 96MB available on the AD102 core). It’ll have a TGP of 450W, the same as the RTX 3090 Ti, and 100W more than the RTX 3090. It should launch with an MSRP of $1,999 sometime next month.
