NVIDIA’s Ada Lovelace GPUs for next-gen GeForce RTX 40 gaming graphics cards will reportedly have a node advantage over AMD’s RDNA 3 Moore’s Law is dead.
NVIDIA Ada Lovelace gaming GPU rumored to use TSMC’s 4N process node, slightly better than AMD’s RDNA 3 GPU
From what we know so far, NVIDIA is expected to use TSMC’s 5nm process node for its Ada Lovelace GPUs to power the next generation of gaming graphics cards, namely the GeForce RTX 40 series. It appears that Moore’s Law has revealed specific nodes in a recent tweet. According to the latest rumors, the NVIDIA Ada Lovelace GPU will be based on the TSMC 4N process node.
Note that I didn’t mention PCIe 5.0 in the Lovelace slide: https://t.co/n7sdIL1Uku
Speaking of “4”, #NVIDIA Lovelace is indeed 4nm!
🤠🎉😋 More on the next broken silicon… https://t.co/jjJ1GSod0j
— Moore’s Law is Dead (@mooreslawisdead) April 24, 2022
Yes, that’s the same TSMC 4N process node that powers Hopper GPUs for the data center HPC market. As for what we know about TSMC’s 4N process node, it’s a revision of their 5nm process (not to be confused with a completely different node, 4nm/N4). The TSMC 4N process node is custom-designed for NVIDIA with a series of optimizations to achieve better power efficiency, performance and small increases in density compared to the normal TSMC 5nm node.
It’s obvious why NVIDIA might choose TSMC’s 4N as a candidate for its next-generation gaming GPU lineup. Upcoming graphics cards will be very power hungry, NVIDIA & the company will optimize them as much as possible using the 4N process node. AMD, on the other hand, will use a mix of TSMC 5nm and 6nm process nodes in its upcoming MCM and RDNA 3 graphics architecture-based monolithic GPUs, and while they do not bring the optimizations that 4N did, they will adopt the MCM approach expected It will be efficient.
So at the end of the day, NVIDIA gets better nodes and AMD offers a better design approach. At the end of the day, these don’t really matter to end users who just want to play games on the best hardware (graphics card) they can get their hands on.
NVIDIA CUDA GPU (rumored) Preliminary:
| graphics processor | TU102 | GA102 | AD102 |
|---|---|---|---|
| Flagship SKUs | RTX 2080 Titanium | RTX 3090 Titanium | RTX 4090? |
| architecture | Turing | ampere | Ada Lovelace |
| process | TSMC 12nm NFF | Samsung 8nm | TSMC 4N? |
| Die size | 754mm2 | 628mm2 | ~600mm2 |
| Graphics Processing Cluster (GPC) | 6 | 7 | 12 |
| Texture Processing Cluster (TPC) | 36 | 42 | 72 |
| Streaming Multiprocessor (SM) | 72 | 84 | 144 |
| CUDA core | 4608 | 10752 | 18432 |
| L2 cache | 6MB | 6MB | 96MB |
| Theoretical TFLOP | 16 TFLOPs | 40 TFLOPs | ~90 TFLOPs? |
| memory type | GDDR6 | GDDR6X | GDDR6X |
| Memory Capacity | 11 GB (2080 Titanium) | 24 GB (3090 Titanium) | 24 GB (4090?) |
| memory speed | 14Gbps | 21Gbps | 24 Gbps? |
| memory bandwidth | 616 GB/sec | 1.008 GB/sec | 1152 GB/sec? |
| memory bus | 384 bit | 384 bit | 384 bit |
| PCIe interface | PCIe Gen 3.0 | PCIe Gen 4.0 | PCIe Gen 4.0 |
| Three Gorges Project | 250W | 350W | 600W? |
| freed | September 2018 | September 20 | 2H 2022 (to be determined) |
