It seems the Nvidia Ada Lovelace is going to be a big GPU, according to rumors. The rumors come from Twitter user kopite7kimi and have been confirmed by the team at 3DCenter. Kopite7kimi is a highly credible source, but since this is a rumor, take everything with a grain of salt.
NVIDIA ADA GPU to be based on the 5nm process architecture with a massive 18432 CUDA core count
Nvidia ADA Lovelace architecture was recently leaked with the release of some preliminary specifications for the upcoming GPU. A recent tweet from Kopite7kimi was followed up by 3DCenter org adding some additional information to the leak.
The AD102 chip is expected to have a massive 18,432 FP32 units and comes with 66 TFlops of computing power. The graphics card runs at 1.8 GHz, which is seemingly great for the AD102 GPU. The speculated AD102 comes with 12 Graphics Processing Clusters, 72 Texture processing clusters, and 144 streaming multiprocessors. The card is already better than the latest Ampere, making the Turing look like a measly architecture.
And a larger cache. It looks like this.
— kopite7kimi (@kopite7kimi) December 28, 2020
The Ampere is already a good architecture with an improvement on the previous Turing Ray-tracing architecture. Kopite7kimi also pointed out the new Lovelace GPU comes with larger cache storage. Both Ampere and Turing architecture came with 6MB of total cache, meaning the ADA GPU will have a full architectural rework. As for the process node, we still can confirm whether Nvidia will stick with Samsung or use TSMC 5nm+ process node.
Nvidia faced a lot of bad yields over at Samsung, slowing down the manufacturing process. The 8nm die yield was already a problem, but it could be a different outcome with the 5nm process node. Using a Samsung’s foundry also has its advantages as TSMC is already all choked up. TSMC fab supplies to prominent industry leaders like AMD, Apple, etc. The fabrication won’t have enough capacity and might slow down Nvidia. It seems Nvidia looks to give another shot at Samsung and its foundry.