More precisely, about its GPU:
Source Kopte7kimi revealed new details about the NVIDIA Ada Lovelace GPU, which will underlie the GeForce RTX 40 series graphics cards.
Kopite7kimi started with GPU configuration. He compared the Ada Lovelace AD102 to other NVIDIA GPUs, specifically the Ampere GA102 and Turing TU102.
The AD102 will receive up to 12 GPCs (Graphics Processing Clusters). This is 70% more than the GA102 with only 7 GPCs.
Each GPU will consist of 6 TPCs and 2 SMs, which matches the configuration of the existing chip. Each SM will contain four sub-cores, which also corresponds to the GA102 GPU.
What has changed is the FP32 and INT32 core configuration. The total number of FP32+INT32 blocks will increase to 192. There will be 18,432 CUDA cores in total.
In terms of cache, the new GPUs will get 192KB of L1 cache per SM block, which is 50% more than the Ampere architecture. There will be a total of 4.5 MB of L1 cache in the AD102. But the cache memory of the second level will be 96 MB – 16 times more than that of the GA102.
Even in the new architecture, the number of rasterization units (ROP) will increase – 32 units per GPC block, or twice as many as now.
Let us know your thought in the comment section, or in our telegram chat.