Please note that this post is tagged as a rumor.
NVIDIA AD102 without PCIe Gen5 support?
Kopite7Kimi, a reliable leaker covering NVIDIA hardware, confirmed today that the next-generation AD102 GPU does not have PCIe Gen5 interface.
For now, it’s just a rumor, but if this was true, then GeForce RTX 40 series would not be a bottleneck for next-gen desktop systems. The AD102 GPU would simply not support any faster interface. Right now, only Intel Alder Lake supports PCI Express 5.0, but around the same time NVIDIA is expected to introduce Ada GPU series, AMD is also preparing its Ryzen 7000 “Raphael” platform, which more than likely already has Gen5 support.
The “Ada Lovelace” architecture which is set to launch by the end of this year was rumored to feature PCIe Gen5 interface support, mainly because it already has the Intel ATX 3.0/PCIe Gen5 spec 16-pin connector unlocking up to 600W of power per cable.
PCIe Gen4
— kopite7kimi (@kopite7kimi) April 24, 2022
The rumor about lack of Gen5 support are especially surprising, considering that back in March this year, NVIDIA introduced Hopper architecture, which was the first GPU architecture to officially support PCIe Gen5. This is how NVIDIA described the new standard for its data-center H100 accelerator:
PCIe Gen 5 provides 128 GB/sec total bandwidth (64 GB/sec in each direction) compared to 64 GB/sec total bandwidth (32GB/sec in each direction) in Gen 4 PCIe. PCIe Gen 5 enables H100 to interface with the highest performing x86 CPUs and SmartNICs or data processing units (DPUs).
Furthermore, Kopite mentioned that the structure of the single-precision (FP32) core configuration might not be as simple as Ampere’s. One should not rule out that the CUDA core count will not change again with the new generation of GeForce RTX series. A matrix of different integer and floating-point FP32 units might increase the number of cores, compute power, or introduce a new core type altogether. It is simply too early to tell.
NVIDIA usually keeps GPU architecture details under a closed lid, even until the very last moment. In fact, when GA102 GPU launched, many board partners had no clue that FP32 core count was to double. We were still finding wrong core count in official data sheets and product descriptions for many days after launch.
| Next-gen Flagship GPU Comparison (RUMORED) | ||
|---|---|---|
| VideoCardz.com | GeForce RTX 4090 – Class | Radeon RX 7900 – Class |
| Fabrication Node | TSMC N5 | TSMC N5/N6 |
| Architecture | NVIDIA Ada | AMD RDNA3 |
| GPU Package | Monolithic | Multi-Chip-Module (MCM) |
| Estimated GPU Size | ~600mm² | ~800mm² |
| Graphics Dies | 1 | 2 GCD + 4 MCD + 1 IOD |
| GPU Mega Clusters | 12 Graphics Processing Clusters (GPC) | 2×3 Shader Engines |
| GPU Super Clusters | 72 Texture Processing Clusters (TPC) | 2×30 RDNA Workgroups (WGP) |
| GPU Clusters | 144 Streaming Multiprocessors (SM) | – |
| FP32 Cores | 18432 CUDA Cores | 15360 Stream Processors |
| GPU Clock | ~ 2.2 GHz | ~ 2.5 GHz |
| Memory Type | 24 GB GDDR6X | TBC GB GDDR6 |
| Memory & Bus | 21 Gbps 384-bit | TBC Gbps 256-bit |
| Cache | 96MB (L2 Cache) | 256 or 512MB Infinity Cache |
| Interface | PCIe Gen4 (?) | TBC |
| Power Consumption | 600W (TDP) | 500W (TBP) |
| Release Date | Q3/Q4 2022 | Q3/Q4 2022 |
Source: @kopite7kimi
