Note that this article is marked as a rumor.
NVIDIA AD102 that does not support PCIe Gen5?
Reliable leaker Kopite7Kimi covering NVIDIA hardware today confirmed that the next-gen AD102 GPU will not have a PCIe Gen5 interface.
For now, that’s just a rumor, but if true, the GeForce RTX 40 series won’t be the bottleneck for next-gen desktop systems. AD102 GPUs simply don’t support any faster interfaces. Currently, only Intel Alder Lake supports PCI Express 5.0, but around the same time NVIDIA is expected to launch its Ada GPU lineup, AMD is also preparing its Ryzen 7000 “Raphael” platform, which will likely already support Gen5.
The “Ada Lovelace” architecture, which is rumored to be launched by the end of this year, supports the PCIe Gen5 interface, mainly because it already has a 16-pin connector of the Intel ATX 3.0/PCIe Gen5 specification, which can unlock up to 600W of power. cable.
Rumors of no Gen5 support are especially surprising since back in March, NVIDIA introduced the Hopper architecture, the first GPU architecture to officially support PCIe Gen5. Here’s how Nvidia describes the new standard for its data center H100 accelerator:
PCIe Gen 5 provides 128 GB/sec of aggregate bandwidth (64 GB/sec in each direction), while PCIe Gen 4 provides 64 GB/sec of aggregate bandwidth (32 GB/sec in each direction). PCIe Gen 5 enables the H100 to interface with the highest performing x86 CPUs and SmartNICs or Data Processing Units (DPUs).
Also, Kopite mentions that the structure of the single-precision (FP32) core configuration may not be as simple as the Ampere’s. It is not excluded that the number of CUDA cores will not change again with the new generation of GeForce RTX series. A matrix of different integer and floating point FP32 units may increase the number of cores, computing power or introduce new core types entirely. It’s too early to tell.
NVIDIA typically keeps GPU architectural details private until the last minute. In fact, when the GA102 GPU was introduced, many board partners were unaware that the FP32 core count would be doubled. Many days after launch, we still find false core counts in official data sheets and product descriptions.
| Next-gen flagship GPUs compared (rumored) | ||
|---|---|---|
| VideoCardz.com | GeForce RTX 4090 – Class | Radeon RX 7900 – Class |
| manufacturing node | TSMC N5 | TSMC N5/N6 |
| architecture | Nvidia Ada | AMD RDNA3 |
| GPU package | Monolithic | Multi-Chip Module (MCM) |
| Estimated GPU size | ~600mm² | ~800mm² |
| Graphic stencil | 1 | 2 GCD + 4 MCD + 1 IOD |
| GPU Large Cluster | 12 Graphics Processing Clusters (GPCs) | 2×3 shader engine |
| GPU super cluster | 72 Texture Processing Clusters (TPCs) | 2×30 RDNA Working Group (WGP) |
| GPU cluster | 144 Streaming Multiprocessor (SM) | – |
| FP32 core | 18432 CUDA cores | 15360 Stream Processor |
| GPU clock | ~ 2.2GHz | ~ 2.5GHz |
| memory type | 24GB GDDR6X | Pending GB GDDR6 |
| memory and bus | 21 Gbps 384 bits | Pending Gbps 256-bit |
| cache | 96MB (level 2 cache) | 256 or 512MB unlimited cache |
| interface | PCIe Gen4 (?) | pending |
| energy consumption | 600W (TDP) | 500W (to be determined) |
| release date | Q3/Q4 2022 | Q3/Q4 2022 |
source: @kopite7kimi
