NVIDIA Blackwell is NVIDIA's latest GPU architecture (2024-2025) — delivering a generational leap in AI training and inference performance with the B200 and GB200 products.

Key Products

• B200 GPU: 208 billion transistors across two dies connected by 10 TB/s NV-HBI chip-to-chip interconnect. 192GB HBM3e at 8 TB/s bandwidth. 9,000 TFLOPS FP4, 4,500 TFLOPS FP8. 1,000W TDP.
• GB200 Grace Blackwell Superchip: One Grace CPU + two B200 GPUs on a single board. 384GB HBM3e total.
• GB200 NVL72 Rack: 72 Blackwell GPUs + 36 Grace CPUs in a single liquid-cooled rack. 1.4 exaFLOPS FP4. 13.5TB HBM3e total. 120kW power draw.
• B100: Lower-power variant for different deployment scenarios.

Architecture Innovations

• Second-Generation Transformer Engine: FP4 precision support — 2x throughput vs FP8 on Hopper
• Fifth-Generation NVLink: 1.8 TB/s bidirectional per GPU (vs 900 GB/s on Hopper)
• NVLink Switch: Enables all 72 GPUs in NVL72 to communicate as a single unified GPU
• Decompression Engine: Hardware-accelerated data decompression for database and analytics workloads
• RAS Engine: Reliability, Availability, Serviceability — dedicated hardware for 24/7 uptime monitoring
• Secure AI: Confidential computing with hardware-based trusted execution environment

Performance vs Hopper (H100)

Training Performance

• GPT-4 class model (1.8T parameters): 4x faster training vs H100 cluster
• GB200 NVL72 rack replaces ~8 DGX H100 systems for equivalent workloads
• Real-time inference for trillion-parameter models becomes practical

Availability and Pricing

• B200 GPU: ~,000-40,000 per unit (estimated)
• GB200 NVL72 Rack: ~-3M per rack
• Cloud: Available on AWS, Azure, GCP, Oracle Cloud (2025)
• Supply: Constrained — HBM3e supply from SK Hynix and Micron is the bottleneck

Customers: Every major AI lab (OpenAI, Anthropic, Google, Meta, xAI), cloud providers (AWS, Azure, GCP, Oracle), and enterprise AI deployments.

Blackwell represents NVIDIA's continued dominance in AI compute — making training 4x faster and inference 30x more efficient than the previous generation, while the NVL72 rack architecture enables trillion-parameter models to run as a single unified system.