GB200 NVL72

The GB200 NVL72 is a Blackwell-generation NVIDIA GPU built on the GB202 architecture, manufactured on a TSMC 4NP process node with 208 billion transistors. Released in 2025, it delivers 4500 TFLOPS of FP16 throughput and 4500 TFLOPS BF16 — the two precision formats most commonly used for transformer model training and inference. FP8 precision is supported at 9000 TFLOPS — roughly 2.0× the FP16 rate — enabling near-doubled throughput for inference workloads that can tolerate reduced numerical precision with calibration. The GB202 architecture represents NVIDIA's approach to balancing compute throughput, memory bandwidth, and power efficiency for data center AI workloads. At 1000W TDP, the GB200 NVL72 sits in the ultra-high-power data center tier (1000W), requiring specialized rack infrastructure with high-density power delivery.

Memory capacity is 192GB of HBM3e with 8000 GB/s bandwidth. This determines which models can run without quantization: approximately 96B parameters at FP16 (2 bytes/param), 192B at INT8 (1 byte/param), or up to 384B parameters at INT4/GGUF quantization (0.5 bytes/param). These figures are theoretical maximums — actual capacity is reduced by KV cache, framework overhead, and activation memory, typically by 10–20% for inference and 30–40% for training. The arithmetic intensity ceiling — the ratio of compute (4500 TFLOPS) to memory bandwidth (8000 GB/s) — is approximately 563 FLOP/byte. Transformer inference is typically memory-bound below this threshold, meaning the 8000 GB/s bandwidth figure is the primary determinant of tokens-per-second for most LLM serving workloads.

Multi-GPU configurations are a first-class use case for the GB200 NVL72. NVLink 5.0 / PCIe 6.0 provides 1800 GB/s of bidirectional NVLink bandwidth between cards, enabling tensor-parallel inference across multiple GPUs with near-linear VRAM scaling. A two-card configuration provides 384GB of pooled VRAM — sufficient for 192B parameter models at FP16 — while a four-card setup reaches 768GB. NVLink's low-latency, high-bandwidth fabric makes all-reduce operations in data-parallel training significantly faster than PCIe-based alternatives, which top out at ~64 GB/s bidirectional for PCIe 5.0 x16.

The primary workloads for the GB200 NVL72 are Frontier LLM training, Trillion-param models, Rack-scale AI. Rack-scale Blackwell system. 72 GPUs + 36 Grace CPUs in one NVL rack. Highest throughput available. Key limitations to factor into your evaluation: Rack-scale only — not available as individual GPUs; Extremely limited availability outside hyperscalers; Requires NVLink Switch infrastructure — very high cost. When comparing this GPU against alternatives at similar price points, the most important metrics are memory bandwidth (for inference throughput), VRAM capacity (for model size), and FP16/BF16 TFLOPS (for training speed). Raw TFLOPS figures can be misleading for inference — a GPU with lower TFLOPS but higher memory bandwidth will often outperform a higher-TFLOPS card on tokens-per-second for memory-bound autoregressive generation.

In the broader GPU market, the GB200 NVL72 is a one-year-old design that remains competitive for most workloads. CUDA compatibility is a significant advantage: the GB200 NVL72 benefits from the full NVIDIA software ecosystem including cuDNN, TensorRT, FlashAttention, and all major inference frameworks (vLLM, TGI, TensorRT-LLM). CUDA's maturity means optimized kernels are available for virtually every model architecture. For cloud rental, availability varies significantly by provider — some specialize in this GPU tier while others may have limited stock. Compare on-demand and spot pricing across providers using the rental comparison table on this page, and factor in region availability if latency is a concern for your inference workload.