©
While this is just a rumor, but if it were true, then GeForce RTX 40 would not be a bottleneck for the next generation of desktop systems. GPU AD102 just won’t support a faster interface. Now only processors Intel Alder Lake support PCI Express 5.0, but at the same time NVIDIAis expected to introduce a series of GPUs “Ada Lovelace”. AMD is also preparing its own platform Ryzen 7000 “Raphael”which most likely already has PCIe Gen5 support.
Architecture “Ada Lovelace”which should be launched by the end of this year, is rumored to have PCIe Gen5 support, mainly because it already has a 16-pin Intel ATX 3.0/PCIe Gen5 connector delivering up to 600W of power per cable.
The rumors regarding the lack of support for Gen5 are especially surprising, given that back in March of this year, NVIDIA introduced the architecture “Hopper”, which became the first architecture to officially support PCIe Gen5. This is how NVIDIA described the new standard for its H100 data center accelerator:
PCIe Gen 5 provides a total bandwidth of 128 GB/s (64 GB/s in each direction) compared to a total bandwidth of 64 GB/s (32 GB/s in each direction) in PCIe Gen 4. PCIe Gen 5 allows the H100 to interoperate with the highest performance x86 processors and SmartNICs or Data Processing Units (DPUs).
Also, source Kopite7Kimi mentioned that single precision (FP32) kernel configuration structure might not be as simple as “Ampere”. It cannot be ruled out that the number of CUDA cores will no longer change with the new generation of the GeForce RTX series. Matrix of various integers and floating point FP32 may increase the number of cores, processing power, or even introduce a new type of core, it’s just too early to tell.
NVIDIA typically keeps GPU architecture details under wraps, even until the very last moment. In fact, when it was launched GA102many board partners had no idea that the number of cores FP32 should double. For quite a long time, the incorrect number of cores was mentioned in the official data sheets and descriptions of different models for many days following the launch.
