NVIDIA RTX Blackwell GPU Architecture: Built for Neural Rendering

Key insights

6 takeaways

01
The GeForce RTX 5090, powered by the GB202 GPU, features a 512-bit memory interface and 32 GB of GDDR7 memory, delivering a peak memory bandwidth of 1792 GB/sec.See it on page 15
02
Blackwell's fifth-generation Tensor Cores introduce FP4 support, which can double AI throughput while halving the memory requirements.See it on page 16
03
The architecture includes a new AI Management Processor based on a RISC-V design to handle GPU context scheduling, reducing reliance on the system CPU.See it on page 25
04
DLSS 4 Multi Frame Generation utilizes AI models to generate up to three additional frames for every traditionally rendered frame; the new frame generation AI model uses 30% less VRAM than the prior frame generation method.See it on page 35
05
The ninth-generation NVENC encoder in Blackwell adds support for 4:2:2 H.264 and HEVC encoding, improving AV1 and HEVC quality by 5% BD-BR PSNR.See it on page 28
06
Blackwell's clock architecture allows for frequency adjustments to dynamic workloads 1000x faster than previous GPU architectures.See it on page 32

Summary

The NVIDIA RTX Blackwell GPU architecture is a hardware design focused on advancing neural rendering, generative AI, and ray tracing capabilities. The architecture introduces several hardware-level innovations, including fifth-generation Tensor Cores with FP4 support, fourth-generation RT Cores, and the integration of GDDR7 memory. These components are designed to support new software features such as DLSS 4, which utilizes multi-frame generation, and neural shaders that allow developers to embed AI models directly into the rendering pipeline. The architecture also features the AI Management Processor, a RISC-V based context scheduler intended to offload task management from the system CPU.

The Blackwell family includes the GB202, GB203, and GB205 GPUs, which power the GeForce RTX 5090, 5080, 5070 Ti, and 5070 graphics cards. These GPUs incorporate Mega Geometry technology to increase geometric detail in ray-traced applications and include new video encoding and decoding support for 4:2:2 chroma formats. Power efficiency is addressed through Max-Q features, including advanced power gating and accelerated frequency switching, which allows clocks to adjust to dynamic workloads 1000x faster than previous architectures. The architecture also introduces support for DisplayPort 2.1b, providing up to 80 Gbps of bandwidth.