The news about users unable to buy the graphics cards from the market isn’t new, and we are still waiting for Nvidia and AMD to fix the supply constraints from their end. The Ampere GPU’s from Nvidia is already fresh in the market but seems to have fallen victim to the supply constraint problem, and it feels like the issue will not be solved anytime soon.
However, NVIDIA is on a fast track to developing Ampere’s successor. The GPU’s are named Lovelace (named after British mathematician Ada Lovelace) and will succeed the Ampere line-up.
According to 3DCenter, NVIDIA AD102, Lovelace’s code-name, will sport 2x the graphics processing clusters (GPCs) of Turing, i.e. 12. This indicates that AD102 may house as many as 72 texture processing clusters (TPCs) and 144 streaming multiprocessors (SMs).
Reports also indicate that the card will have a total of 18,432 CUDA cores — a 71% increase compared to Ampere’s GA102 die. The clock speed is assumed to at the base of about 1.8 GHz, and could theoretically offer nearly 66 TFLOPs of FP32 single-precision performance.
But we still don’t know whether Nvidia is planning to go with Samsung or TSMC for its nm process nodes for Lovelace. We also know that Nvidia has already pre-booked the production of TSMC’s 5nm nodes for 2021, which makes Samsungs chances pretty low in the competition.
| Lovelace AD102 | Ampere GA102 | Turing TU102 | |
| Process | Samsung 5nm (?) | Samung 8 nm | TSMC 12 nm |
| Graphics Processing Clusters | 12 | 7 | 6 |
| Texture Processing Clusters | 72 | 42 | 36 |
| Streaming Multiprocessors | 144 | 84 | 72 |
| CUDA Cores | 18,432 | 10,752 | 4,608 |
| FP32 at 1.8 GHz | 66.4 TFLOPs | 38.7 TFLOPs | 16.6 TFLOPs |
| Memory Bus | 384-bit (?) | 384-bit | 384-bit |
| Memory Type | GDDR6X (?) | GDDR6X | GDDR6 |
