NVIDIA Blackwell is a next-generation GPU architecture unveiled by NVIDIA in 2024, designed as the 'engine of AI factories' for training and inference of large models. Blackwell chips (including the B200) contain 208 billion transistors and are manufactured on a custom TSMC 4NP process. They use a dual-die design: two reticle-limited dies are connected by a 10 TB/s chip-to-chip interconnect (NV-HBI) and operate as a single, coherent GPU.
Key innovations include the second-generation Transformer Engine with 4-bit floating point (FP4) support - doubling the performance and size of next-generation models - and fifth-generation NVLink (up to 1.8 TB/s per GPU) for multi-chip interconnect. Blackwell adds a dedicated RAS engine (reliability, availability, serviceability), a decompression engine for data analytics, and confidential computing (TEE-I/O) for security. The B200 chip offers 192 GB of HBM3e memory at about 8 TB/s bandwidth and up to about 20 petaFLOPS of FP4.
Blackwell is the foundation of a broad NVIDIA product family: the B200/B100 chips, the GB200 superchip (2 Blackwell GPUs + a Grace CPU) and the GB200 NVL72 and GB300 NVL72 rack-scale systems. The architecture's successor is NVIDIA Rubin (Vera Rubin).
AI Accelerator
Which group NVIDIA Blackwell belongs to and how it is built
An AI Accelerator is a specialized hardware component designed for efficient execution of artificial intelligence computations, particularly neural network inference, computer vision processing, and sensor data analysis. In robotics, AI accelerators are used to run perception models, object recognition, image segmentation, planning, and other tasks that require high computational throughput under constrained power budgets. They may take the form of dedicated NPU, TPU, VPU, or GPU chips, or specialized embedded modules.
A data-center AI accelerator card is a design class describing the construction of high-performance compute processors (GPUs/accelerators) intended for mounting in data-center servers. It is characterised by: an SXM form factor (a module soldered onto an HGX/DGX baseboard) or a dual-slot PCIe card; high-bandwidth memory (HBM2e/HBM3/HBM3e) integrated on-package; dedicated GPU-to-GPU interconnects (NVLink, Infinity Fabric) with hundreds of GB/s of bandwidth; high TDP (350–1000 W) requiring air or liquid cooling; support for virtualisation/partitioning (MIG) and low-precision compute formats (FP8/FP16/BF16/INT8). The class includes designs such as NVIDIA H100/H200/A100, AMD Instinct MI300, Google TPU, and Intel Gaudi. It describes physical construction and configuration, not the functional role (which is given by the component type "AI Accelerator").
Basic physical properties of NVIDIA Blackwell — dimensions, weight and materials
Other hardware parts related to NVIDIA Blackwell
Rack-scale, liquid-cooled computing system built around 72 NVIDIA Blackwell (B200) GPUs and 36 NVIDIA Grace CPUs, designed for training and inference of large AI models.
Rack-scale, liquid-cooled computing system built around 72 NVIDIA Blackwell Ultra (B300) GPUs and 36 NVIDIA Grace CPUs, designed for large AI model inference and training.
A data-center AI accelerator based on the NVIDIA Hopper architecture (2023): the first GPU with HBM3e memory — 141 GB at 4.8 TB/s. Same compute as the H100 but nearly double the memory capacity and bandwidth for training and inference of large language models.
A data-center AI accelerator based on the NVIDIA Hopper architecture (2022): 80 GB HBM3, 700 W, 4th-generation Tensor Cores with FP8 and Transformer Engine. The standard GPU for LLM training and hyperscale AI inference.
NVIDIA's next-generation AI platform: Rubin GPU (HBM4) + Vera CPU (88 Olympus cores). The Vera Rubin NVL144 rack: 144 Rubin GPUs, 36 Vera CPUs, up to 3.6 EF FP4 (inference) and 1.2 EF FP8 (training). Availability: 2H 2026.
