The Blackwell Series is NVIDIA's next-generation graphics processing unit (GPU) computing architecture designed specifically for hyperscale artificial intelligence (AI), deep learning, and high-performance computing (HPC). Named after mathematician David Blackwell, this architecture succeeds the Hopper architecture to power massive large language models (LLMs) and generative AI workloads at a fraction of the energy consumption.
Generative AI Focus: Engineered specifically to handle trillion-parameter AI models efficiently.
Dual-Die Design: Combines two distinct compute dies into a unified chip via an ultra-high-speed interconnect.
Second-Generation Transformer Engine: Utilizes microscopic FP4 precision formatting to double processing speeds for AI training and inference.
De-risking Efficiency: Reduces energy consumption and operational costs up to 25 times compared to previous architectures.
NVIDIA structures its technology roadmap by alternating focus between consumer graphics and enterprise data center architectures.
Ampere (2020): Introduced unified architecture for both gaming (RTX 30-series) and data centers (A100), standardizing AI tensor cores.
Hopper (2022): Shifted to a dedicated data center split, optimization for LLM training with the H100 GPU using the first-generation Transformer Engine.
Blackwell (2026): Represents a shift from monolithic silicon design to multi-die packaging, optimizing data pipelines for trillion-parameter AI systems.
Blackwell moves away from traditional single-chip manufacturing limits by implementing a multi-die approach. Two fully capable silicon dies are manufactured separately and bound together using a high-bandwidth chiplet interconnect running at 10 terabytes per second (TB/s).
To software and compilers, this dual-die configuration acts as a single unified GPU. This prevents data bottlenecks across the chip. Data flows into massive High Bandwidth Memory (HBM3e), which is then processed through the Second-Generation Transformer Engine. This custom engine dynamically scales down numerical precision to 4-bit floating point format (FP4) during compute phases where full precision is mathematically unnecessary, doubling processing throughput instantly without losing model accuracy.
A high-speed communication interface that allows up to 576 individual GPUs to talk to one another simultaneously, delivering 1.8 TB/s of bidirectional bandwidth per GPU to treat massive server clusters as a single computing entity.
A hardware-based accelerator that offloads data decompression tasks from the CPU directly onto the GPU. It speeds up data analytics pipelines and accelerates data transfer from storage to memory.
Advanced hardware-based confidential computing capabilities that protect sensitive AI models, proprietary data, and cryptographic keys from unauthorized access during training or inference phases.
The Blackwell architecture is deployed across several hardware form factors tailored for distinct data center requirements:
NVIDIA B100 / B200 GPUs: Standard discrete server accelerators designed to drop into existing air-cooled or liquid-cooled data center server architectures.
NVIDIA GB200 NVL72: A complete liquid-cooled rack-scale system that links 36 Grace CPUs and 72 Blackwell GPUs together via NVLink switches, acting as a singular giant GPU cluster.
Feature | NVIDIA Hopper (H100) | NVIDIA Blackwell (B200) |
|---|---|---|
Silicon Design | Monolithic (Single Die) | Multi-Die (Dual-Chiplet) |
Transistor Count | 80 Billion | 208 Billion |
Lowest AI Precision | FP8 (8-bit Floating Point) | FP4 (4-bit Floating Point) |
Interconnect Speed | 900 GB/s (NVLink 4) | 1.8 TB/s (NVLink 5) |
Memory Type | HBM3 | HBM3e |
While Blackwell significantly reduces energy consumption per calculation, the massive scale of these chips requires specialized liquid-cooling infrastructure in modern data centers. Additionally, extracting the maximum performance benefit requires specialized software compilation to utilize the new FP4 low-precision data types effectively.
Trillion-Parameter Model Inference: Powering real-time responses from complex, multimodal generative AI systems.
Quantum Computing Simulations: Modeling complex quantum mechanics and chemical molecular structures on classic supercomputers.
Automated Scientific Research: Accelerating genomic sequencing analysis and drug discovery pipelines through rapid deep learning data parsing.
Tensor Core: A specialized execution unit inside a GPU designed specifically to accelerate matrix multiplication mathematics used in AI.
HBM3e: High Bandwidth Memory 3 Extended, a stacked memory architecture providing extreme data transmission speeds directly adjacent to the processor.
Chiplet: An integrated circuit block containing a specific subset of functionality, designed to combine with other chiplets to form a larger processor.
Learn how AMD Fluid Motion Frames (AFMF) works, its hardware requirements, pros, cons, and how it differs from FSR 3 in this complete hardware glossary.
Learn how RT Cores accelerate real-time ray tracing, improve gaming visuals, and offload critical BVH calculations from standard GPU shaders.
Learn about GDDR7 graphics memory, including its PAM3 signaling technology, technical specifications, and how it advances GPU performance.
Discover what Render Output Units (ROPs) do in your graphics card. Learn how they process pixels, impact 4K gaming, and influence GPU performance.
Learn about the AMD Radeon Pro Series. Discover how these certified workstation GPUs power CAD, 3D rendering, and professional creator workloads.