NVIDIA seems to have taken notice of AMD’s ambitious plans with its next-gen RDNA 3 graphics architecture, and devised plans of tackling it as well. Team Red is planning to adopt a chiplet design with up to seven dies for the Radeon RX 7900 XT (Navi 31), allowing for a remarkable 2-2.5x increase in raw graphics performance. As reported a short while back, AMD has significantly changed the WGP and the number of active threads per WGP with RDNA 3.
I must clarify that the current AD102 is NOT the original AD102. Ada Lovelace is no longer a simple Ampere refresh, although it was like this in the beginning. (zorafestival.org)
— kopite7kimi (@kopite7kimi) April 29, 2022
NVIDIA, on the other hand, had a simple and straightforward plan: Take Ampere, port it to TSMC’s 5nm process, and cram as many transistors as possible onto an 800+mm^2 die. That design seems to have been scrapped and replaced with a much more potent one. According to Kopite7kimi, the foremost source on NVIDIA’s future GPU, the AD102 core has been updated and can no longer be compared to the original die.
| 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? |
It’s unclear what part of the Lovelace architecture has been updated, but you can now expect the RTX 4090 and the Radeon RX 7900 XT to offer roughly the same level of performance. Recent rumors predict an incredible performance target of 100 TFLOPs for the two flagships. This will be achieved using a combination of increased core and boost clocks with unreasonably high TGPs to boot.
Previous coverage:
NVIDIA’s GeForce RTX 4090 is going to come with 24GB of 24Gbps GDDR6X memory paired with the AD102-300 die and a TGP of 600W. This scoop was shared by @kopite7kimi, one of the most reliable sources on NVIDIA’s GPUs. According to Kimi, the RTX 4080 and 4090 will be based on the AD102 die and 24Gbps GDDR6X memory with the latter featuring a cutdown bus and a few SMs disabled.
The GeForce RTX 4080 is expected to feature 20GB of GDDR6X memory across a 320-bit bus. This GPU will feature around 14,000 FP32 cores, a notable cut from the 15-16K shaders on the RTX 4090. The latter will be paired with a 384-bit bus instead. As for the RTX 4070, we’re looking at 12GB of GDDR6X memory across a 256-bit bus. It will be based on the AD104-400 die and pack up to 10,000 FP32 cores. Finally, the RTX 4070 and 4080 will have a TGP of 400W and 500W, respectively.
