NVIDIA’s Blackwell architecture marks a significant departure from its predecessor, Hopper, by introducing a dual-die design and native support for 4-bit floating-point precision. Announced in March 2024, this platform is now the backbone of high-performance AI infrastructure, powering everything from single-server deployments to massive rack-scale systems. For technical readers, understanding the nuances between the B200, B300, and their superchip counterparts is essential for optimizing modern AI workloads.
The Dual-Die Architectural Shift
The most visible change in Blackwell is its packaging. While Hopper utilized a single die with 80 billion transistors, Blackwell splits 208 billion transistors across two dies, each near the reticle limit of TSMC’s 4NP process. These dies are connected via a high-bandwidth interface called NV-HBI, which provides 10 TB/s of bandwidth. NVIDIA treats this assembly as a unified single GPU, allowing it to be programmed and scheduled as one device rather than two separate accelerators. This approach allows for a significant increase in transistor count within a single process generation, albeit with a higher power budget.
FP4 Precision and Transformer Engine
Micro-Tensor Scaling
Blackwell’s second-generation Transformer Engine extends precision support to 4-bit floating point (FP4). This is achieved through micro-tensor scaling, where small blocks of values receive individual scale factors to maintain accuracy. NVIDIA’s NVFP4 format applies two levels of scaling, offering accuracy close to FP8 while reducing memory usage by approximately 1.8 times. This efficiency is critical for inference, where memory bandwidth often becomes the bottleneck.
NVIDIA says this gives accuracy close to FP8 while cutting memory use by about 1.8 times compared with FP8.
On paper, the throughput gains are substantial. While Hopper delivers 2 petaFLOPS of dense FP8, Blackwell reaches 5 petaFLOPS of dense FP8 and 10 petaFLOPS of dense NVFP4. The Blackwell Ultra variant, known as B300, further raises dense NVFP4 throughput to 15 petaFLOPS. For a 70-billion parameter model, storing weights in a 4-bit format requires only about 35 GB, leaving significant headroom on a B200 for the key-value cache.
Memory, Power, and Networking
The standard B200 GPU features up to 192 GB of HBM3E memory with 8 TB/s of bandwidth and a maximum power draw of 1,200 W. The B300 (Blackwell Ultra) increases this to 288 GB of HBM3E while maintaining the same bandwidth, with a power ceiling of 1,400 W. Both generations benefit from NVLink 5, which doubles per-GPU bandwidth to 1,800 GB/s. This allows for larger scale-up domains, with the NVL72 rack connecting 72 GPUs in a single NVLink domain, providing 130 TB/s of aggregate bandwidth.
Grace Blackwell Superchips
The GB200 and GB300 superchips integrate two Blackwell GPUs with one Arm-based Grace CPU. This configuration is linked via NVLink chip-to-chip (NVLink-C2C) at 900 GB/s, enabling coherent memory access that far exceeds PCIe bandwidth. These superchips are the building blocks for the GB200 NVL72 and GB300 NVL72 racks. For buyers, this means selecting a rack or a partition of one, rather than a standalone server. The GB300 NVL72 is rated at approximately 1.1 exaFLOPS of dense FP4, making it a formidable tool for large-scale training and inference.
Practical Implications for AI Workloads
For teams deploying AI models, the primary benefits of Blackwell are increased memory per GPU and higher bandwidth. This determines how many GPUs are needed for a specific model and whether the facility can support the power requirements. While vendor benchmarks claim up to 30 times faster real-time inference for large mixture-of-experts models, per-GPU gains on smaller models are more modest. However, the shift to FP4 and the expanded NVLink domain provide a clear path for scaling complex reasoning workloads.
What to Watch Next
As of late 2026, Blackwell remains the mainstream generation, but NVIDIA has announced that its successor, Vera Rubin, is ramping into full production. Vera Rubin will feature sixth-generation NVLink with 3,000 GB/s per GPU. For now, Blackwell and Blackwell Ultra offer the best balance of performance and availability for high-volume AI inference and training. Readers should monitor the transition to Rubin systems, which are expected to begin shipping in the fall of 2026, as this will further reshape the landscape of AI hardware.