For much of the past decade, progress in data center GPUs has been measured by how quickly they could train ever-larger models. With Blackwell Ultra, NVIDIA shifts that center of gravity. This architecture is explicitly optimized for inference, reasoning, and test-time compute—phases where models consume vastly more tokens than during training and where memory behavior, latency, and power efficiency matter as much as raw peak FLOPS.
Within this shift, two products define a new baseline for AI infrastructure: the B300 GPU and the GB300 NVL72 rack-scale system. Together, they turn Blackwell Ultra from an incremental evolution into a reference design for how large-scale enterprise and sovereign AI systems will be built and operated.
🧠 What “Ultra” Changes Inside Blackwell #
Blackwell Ultra continues the Blackwell lineage, but with design decisions that clearly favor inference-heavy and reasoning-centric workloads.
At the silicon level, Blackwell Ultra delivers roughly 1.5× higher performance than standard Blackwell. Native NVFP4 precision effectively doubles usable compute density while reducing memory footprint, maintaining model accuracy for transformer-based workloads.
One of the most consequential changes is a doubling of attention-layer acceleration. Attention now dominates inference cost for long-context and reasoning-heavy models. While matrix multiplication throughput has scaled rapidly across generations, the Special Functions Unit (SFU)—responsible for exponentials and other transcendental math used in softmax—has historically lagged behind. Blackwell Ultra addresses this imbalance with enhanced SFU capability, translating directly into faster attention execution and lower inference cost for transformer models.
Memory capacity also increases dramatically. Each Blackwell Ultra GPU integrates 288 GB of HBM3e, enabled by 12-high HBM stacks rather than the 8-high configurations of earlier designs. This leap is critical for large language models, retrieval-augmented generation pipelines, and mixture-of-experts architectures that require large resident weight sets, KV caches, and activations. These changes reflect a strategic focus on steady-state inference economics rather than episodic training runs.
The die composition reinforces this intent. Blackwell Ultra retains a dual-reticle design, allocating more silicon area to tensor cores and memory paths. Traditional FP64 compute—long a marker of HPC capability—is intentionally deemphasized. Ultra is not designed as a general-purpose scientific accelerator; it is purpose-built for industrial-scale AI.
🧩 The Blackwell Ultra Product Stack: B300 and GB300 NVL72 #
B300 GPU #
The B300 is the fundamental compute building block of Blackwell Ultra. It is a memory-dense, inference-weighted processor designed for modern LLMs, agentic systems, and long-context reasoning workloads.
Key characteristics include a dual-reticle Blackwell Ultra design, significantly increased NVFP4 and tensor core throughput, and 288 GB of HBM3e per GPU. B300 deliberately trades traditional FP64 capability for maximum efficiency in low-precision compute, fast attention execution, and sustained memory bandwidth.
In practice, B300 typically appears in 8-GPU DGX or HGX systems, forming the basic scheduling and deployment unit around which platform teams design inference infrastructure.
GB300 NVL72 #
The GB300 NVL72 represents a shift from node-centric thinking to rack-scale design. It integrates 72 B300-class GPUs, 36 Grace CPUs, and a next-generation NVLink fabric into a single, coherent accelerator domain.
With more than 20 TB of aggregate HBM and massive NVLink bandwidth, GB300 behaves less like a cluster and more like a monolithic super-accelerator. While B300 optimizes the GPU itself—memory capacity, attention throughput, low-precision efficiency—the GB300 optimizes the system: power delivery, cooling, interconnect topology, and rack-scale coherence.
This system-level focus makes GB300 the deployment primitive for sovereign AI installations, enterprise inference factories, and next-generation AI clouds.
📊 B300 vs. GB300 NVL72 Comparison #
| Feature | NVIDIA B300 GPU | NVIDIA GB300 NVL72 |
|---|---|---|
| Role | Core GPU compute building block | Rack-scale unified accelerator system |
| Architecture | Dual-reticle Blackwell Ultra GPU with NV-HBI links | 72× Blackwell Ultra GPUs + 36 Grace CPUs via NVLink Switch |
| Memory | 288 GB HBM3e per GPU | ~20+ TB total HBM |
| Compute Focus | NVFP4/FP8 inference, reasoning, attention acceleration | Rack-scale reasoning, long-context LLMs, agentic workloads |
| Primary Bottleneck Addressed | GPU memory capacity and attention throughput | Interconnect coherence, power and cooling density |
| Deployment Form | 8-GPU DGX/HGX systems | Self-contained, liquid-cooled 120+ kW rack |
| Ideal Use Cases | High-throughput inference, MoE, test-time scaling | Multi-trillion-parameter models, massive concurrency |
⚙️ Hardware and Software Co-Design #
The gains of Blackwell Ultra are not purely architectural. NVIDIA has emphasized NVFP4 quantization workflows, new parallelism strategies, and improved sharding models tuned for NVLink 5 fabrics. Ultra’s advantage emerges from tight coupling between hardware specialization and software capable of saturating it.
A key signal is where performance improvements are largest. The most consistent gains appear in attention-bound workloads, not dense matrix multiplication. This reinforces the core design intent: Blackwell Ultra is optimized for inference and reasoning economics rather than brute-force FP8 training alone.
🔮 Where Blackwell Ultra Points the Industry #
Low-Precision Compute Over FP64 #
Blackwell Ultra strongly prioritizes NVFP4 and FP8 tensor formats. Workloads dependent on FP64 precision—such as many traditional HPC and physics simulations—will see less relative benefit. This is a deliberate signal that NVIDIA is doubling down on AI inference and reasoning rather than pursuing a universal accelerator model.
Memory Density and Physical Constraints #
With 288 GB of HBM per GPU and dozens of GPUs per rack, memory capacity is no longer the primary bottleneck. Power delivery and thermal density become first-order constraints. GB300 NVL72 systems typically require liquid cooling and specialized power infrastructure, reshaping how data centers are planned and provisioned.
A New Operating Model for AI Infrastructure #
Blackwell Ultra represents more than a performance uplift. It redefines what a GPU is expected to do: memory-dense, low-precision tensor compute at massive scale, optimized for inference, reasoning, and rack-level coherence. For enterprises, cloud providers, and research institutions planning multi-year AI strategies, this shift changes how clusters are designed, scheduled, cooled, and powered.
With Blackwell Ultra, organizations are no longer simply provisioning GPUs. They are building AI factories, optimized end-to-end for accelerated computing and the next generation of AI workloads.