Next-generation NVIDIA GeForce RTX Lovelace GPU rumors bring up to 18,432 CUDA cores for monster performance

GeForce RTX 3090
We barely had a chance to familiarize ourselves with Ampere, with NVIDIA having released its GeForce RTX 30 series relatively recently. Even so, the chatter of the drinking fountain has already changed to what comes next. Rumors say there is a 5-nanometer “Lovelace” GPU architecture that will eventually supplant the Ampere, and it has been speculated that it could arrive with up to 18,432 CUDA cores.

Let’s put that in perspective, okay? The current generation GeForce RTX 3090 is the fastest consumer card that NVIDIA has ever released. It is built around a GA102 Ampere GPU with 72 streaming multiprocessors and 10,496 CUDA cores, along with 328 texture mapping units, 112 rendering output units, 328 Tensor cores and 82 dedicated beam tracking cores.

You can check out our review of the GeForce RTX 3090 to see what kind of numbers it can display in games, but simply, it is the “fastest consumer GPU on the market today, without exception”, as we wrote. Now imagine almost doubling the number of CUDA cores. For additional reference, here is the current line from NVIDIA …
  • GeForce RTX 3090: 72 SMs, 10,496 CUDA cores
  • GeForce RTX 3080: 68 SMs, 8,704 CUDA cores
  • GeForce RTX 3070: 46 SMs, 5,888 CUDA cores
  • GeForce RTX 3060 Ti: 38 SMs, 4,864 CUDA cores
Returning to the GPU Lovelace rumors, she first appeared on a supposed script last week, by prominent leaker and Twitter user @ kopite7kimi. According to the leak, NVIDIA initially planned to launch a GPU called Hopper before Lovelace, but things have apparently changed, and Lovelace will debut first. Hopper will be released later and can potentially be relegated strictly to the high performance computing (HPC) market.

The same leak now suggested that Lovelace could wield an 18,432 CUDA core. Take a look…

NVIDIA Lovelace CUDA Colors

It seems more speculation than insider information, but given the history of the spill, it may be speculation based on privileged information. Unfortunately @ kopite7kimi does not extend over the assumed 12 * 6 structure, but it is interesting nonetheless.

This is equivalent to an increase in graphics processing clusters (GPCs) to 12, against a GA102 Ampere GPU full of 7 GPCs. If necessary, Lovelace could potentially accommodate 72 texture processing clusters (TPCs) and 144 streaming multiprocessors, giving way to an astonishing 18,432 CUDA cores.

That would be an absolute monster, if the information is accurate. And who really knows at this early stage. Remember that Turing was first released in September 2018, while Ampere was released two years later. So this is an architecture that we may not see until the end of 2022.

Source