We’ve barely had a chance to get to know the Ampere, with NVIDIA recently launching the GeForce RTX 30 series. Nevertheless, the water cooler talk has already turned to what comes next. Rumor has it that a 5-nanometer “Lovelace” GPU architecture is threatening to eventually replace the Ampere, and it has been speculated that it could come with up to 18,432 CUDA cores.
Let’s put it in perspective, shall we? The current generation GeForce RTX 3090 is the fastest consumer card NVIDIA has ever launched. It is built around a GA1
02 Ampere GPU with 72 streaming multiprocessors and 10,496 CUDA cores, along with 328 structure mapping units, 112 reproducing output units, 328 Tensor cores and 82 dedicated beam cores.
You can check out the GeForce RTX 3090 review to see what types of numbers it can put into play, but it’s simply the “fastest consumer GPU on the market at the moment, without any”, as we wrote. Imagine almost doubling the number of CUDA cores. For extra reference, here is NVIDIA’s current lineup …
- GeForce RTX 3090: 72 SM, 10,496 CUDA cores
- GeForce RTX 3080: 68 SM, 8,704 CUDA cores
- GeForce RTX 3070: 46 SM, 5,888 CUDA cores
- GeForce RTX 3060 Ti: 38 SMs, 4864 CUDA cores
When we got back to the rumored Lovelace GPU, this first appeared in an alleged roadmap last week, by prominent leaker and Twitter user @ kopite7kimi. According to the leaker, NVIDIA had originally planned to launch a GPU called Hopper before Lovelace, but things have apparently changed, and Lovelace will debut first. Hopper comes out later, and can potentially be referred strictly to the HPC (high performance computing) market.
The same leak has now indicated that Lovelace could use 18,432 CUDA cores. Take a look…
It sounds more like speculation than inside information, but given the leaker’s track record, it can speculate based on inside information. Unfortunately @ kopite7kimi does not elaborate on the supposed 12 * 6 structure, but it is still interesting.
What this represents is a bump in graphics processing clusters (GPCs) to 12, against a full GA102 Ampere GPU having 7 GPCs. If accurate, Lovelace could potentially accommodate 72 texture processing clusters (TPCs) and 144 streaming multiprocessors, giving way to as many as 18,432 CUDA cores.
It would be an absolute monster if the information is accurate. And who really knows at this early stage. Remember that Turing was first launched in September 2018, while Ampere came out two years later. So this is an architecture that we may not see until the latter part of 2022.